Wind power grid-connected regulation system based on multivariate data analysis

Abstract

The present application relates to the field of power system management, and provides a wind power grid-connected regulation system based on multivariate data analysis, which first obtains real-time data of wind farms and power grids, then uses time series encoding and decoding technology to predict future short-term power fluctuations of wind farms, and combines the real-time data of the power grid to evaluate the expected frequency impact, and then determines the frequency risk level based on the power fluctuation prediction and frequency impact evaluation results, and then inputs the prediction results, evaluation results, risk level and real-time data of the power grid into the active frequency support strategy generation module to generate control instruction parameters, and finally sends the control instruction parameters to the wind farm coordination controller for adjusting the physical action of the wind turbine to achieve active frequency support. In this way, it is conducive to improving the efficiency of wind power grid connection in participating in frequency regulation, and can effectively provide technical guarantees for the stable operation of a high proportion of new energy power grids.

Description

Technical Field

[0001] The present application relates to the field of power system management, and more specifically, to a wind power grid-connected regulation system based on multivariate data analysis. Background Art

[0002] Driven by the global transformation of the energy structure, wind power, as a key renewable energy source, is experiencing unprecedented growth, with its penetration rate in power systems continuously increasing. However, due to the intermittent and fluctuating nature of wind power, this poses a serious challenge to the safe and stable operation of power systems, particularly the impact on grid frequency stability. Large-scale wind power integration can cause grid frequency fluctuations, which, if improperly regulated, could threaten grid security. Traditional power systems rely on the inertia and primary frequency regulation capabilities of synchronous generators to maintain frequency stability. However, modern wind turbines using power electronic converter interfaces have decoupled rotational inertia from the grid, resulting in limited response to frequency changes and requiring active control strategies to participate in system frequency regulation.

[0003] To address these challenges, there is an urgent need to develop an advanced wind power grid-connected regulation system. This system should accurately predict wind power fluctuations based on real-time data and intelligently participate in grid frequency regulation to improve grid friendliness. Although existing wind power grid-connected control strategies and systems attempt to address some of these issues, they still have several shortcomings. First, data utilization is insufficient. They often focus solely on wind farm data (such as wind speed and power) while ignoring real-time dynamic information on the grid side (such as system frequency, frequency change rate, tie-line flow, and load levels). This results in regulation decisions that lack an accurate understanding of the grid's actual needs. Second, the prediction accuracy of short-term, drastic fluctuations in wind power is low, making it difficult to support the formulation of forward-looking, refined frequency support strategies. Finally, control strategies are often simplistic or static, failing to dynamically adjust control parameters (such as virtual inertia, reserve capacity, and response rate) based on predicted fluctuation characteristics, assessed frequency impacts, and real-time grid conditions. This makes it difficult to achieve optimal frequency support.

[0004] Therefore, an optimized wind power grid-connected regulation scheme is desired. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a wind power grid-connected regulation system based on multivariate data analysis.

[0006] According to one aspect of the present application, a wind power grid-connected regulation system based on multivariate data analysis is provided, comprising:

[0007] A multivariate data acquisition module, configured to acquire multivariate data, wherein the multivariate data includes real-time data of wind farms and real-time data of power grids;

[0008] A short-term power fluctuation prediction module is used to perform sequence encoding and decoding on the time series of the real-time data of the wind farm to obtain a prediction result of future short-term power fluctuations;

[0009] an expected frequency impact analysis module, configured to input the future short-term power fluctuation prediction result and the real-time power grid data into a frequency response analysis model to obtain an expected frequency impact assessment result;

[0010] a level generating module, configured to determine a frequency risk level based on the future short-term power fluctuation prediction result and the expected frequency impact assessment result;

[0011] a control instruction parameter generation module, configured to input the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level, and the real-time grid data into an active frequency support strategy generation module to obtain control instruction parameters, wherein the control instruction parameters include a virtual inertia coefficient, a reserved standby percentage, a power change rate limit value, and a target active power adjustment value;

[0012] The wind turbine adjustment module is configured to send the control instruction parameters to a wind farm coordination controller, and the wind farm coordination controller adjusts the physical actions of the wind turbine based on the control instruction parameters.

[0013] Compared with the existing technology, the wind power grid-connected regulation system based on multivariate data analysis provided by this application first obtains real-time data of the wind farm and the power grid, then uses time series encoding and decoding technology to predict the future short-term power fluctuations of the wind farm, and combines the real-time data of the power grid to evaluate the expected frequency impact, and then determines the frequency risk level based on the power fluctuation prediction and frequency impact evaluation results, and then inputs the prediction results, evaluation results, risk level and real-time data of the power grid into the active frequency support strategy generation module to generate control instruction parameters, and finally sends the control instruction parameters to the wind farm coordination controller for adjusting the physical action of the wind turbine to achieve active frequency support. In this way, it is conducive to improving the efficiency of wind power grid connection in participating in frequency regulation, and can effectively provide technical guarantees for the stable operation of a high proportion of new energy power grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 14 is a system block diagram of a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application.

[0016] Figure 2 4 is a block diagram of a short-term power fluctuation prediction module in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application.

[0017] Figure 3 4 is a block diagram of a wind speed-power time series reasoning unit in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application.

[0018] Figure 4 4 is a block diagram of an implicit fine-grained association subunit in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0020] Based on this, this application proposes a wind power grid-connected regulation system based on multivariate data analysis. Figure 1 FIG is a system block diagram of a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application. Figure 1 As shown, in the wind power grid-connected regulation system 100 based on multivariate data analysis, it includes: a multivariate data acquisition module 110 for acquiring multivariate data, wherein the multivariate data includes real-time data of wind farms and real-time data of power grids; a short-term power fluctuation prediction module 120 for performing sequence encoding and decoding on the time series of the real-time data of the wind farms to obtain a future short-term power fluctuation prediction result; an expected frequency impact analysis module 130 for inputting the future short-term power fluctuation prediction result and the real-time data of the power grid into a frequency response analysis model to obtain an expected frequency impact assessment result; a level generation module 140 for generating an expected frequency impact assessment result based on the future short-term power fluctuation prediction result. and the expected frequency impact assessment result, determine the frequency risk level; a control instruction parameter generation module 150, used to input the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid into the active frequency support strategy generation module to obtain control instruction parameters, the control instruction parameters include virtual inertia coefficient, reserved standby percentage, power change rate limit value and target active power adjustment value; a wind turbine adjustment module 160, used to send the control instruction parameters to the wind farm coordination controller, and the wind farm coordination controller adjusts the physical action of the wind turbine based on the control instruction parameters.

[0021] Specifically, by integrating real-time wind farm data (such as wind speed and power) with dynamic grid data (such as frequency change rate and load), this system achieves collaborative perception of multi-dimensional data, addressing the shortcomings of traditional methods in ignoring grid-side demand information and making regulation decisions more aligned with actual operating conditions. Secondly, it performs short-term, high-precision forecasts of wind farm power fluctuations. Combined with frequency response analysis models, it quantifies their impact on grid frequency, allowing for a preemptive assessment of frequency risk levels and providing a data foundation for proactive control. Finally, by dynamically generating control parameters such as the virtual inertia coefficient and reserve reserve percentage, and adaptively adjusting strategies based on real-time grid conditions and forecast results, it addresses the inability of traditional static control models to adapt to high-frequency fluctuations. For example, when a short-term power dip is predicted and the frequency risk level is high, the system can dynamically increase the virtual inertia coefficient and reserve capacity, while simultaneously limiting the power change rate to smooth the regulation process, thereby simultaneously optimizing inertia compensation and power balancing capabilities. Through a data-driven closed-loop control mechanism, this system systematically improves the efficiency of wind power grid-connected frequency regulation in terms of forecast accuracy, response speed, and dynamic parameter optimization, providing technical support for the stable operation of power grids with a high proportion of renewable energy.

[0022] In an embodiment of the present application, the multivariate data acquisition module 110 is used to acquire multivariate data, and the multivariate data includes real-time data of wind farms and real-time data of power grids. Specifically, the real-time data of wind farms include wind speed values ​​and power values; the real-time data of power grids include system frequency of the grid connection point, frequency change rate, tie line power and regional load. It should be understood that the real-time data of wind farms specifically include key physical quantities that reflect the operating status of wind farms and the energy conversion process, wherein the wind speed value refers to the air flow velocity data collected in real time by a wind speed sensor installed at the height of the wind turbine hub in the wind farm, which is used to characterize the original driving conditions of wind energy input; the power value specifically refers to the active power value measured in real time at the wind farm grid connection point, reflecting the actual output capacity of the wind turbine to convert wind energy into electrical energy and inject it into the power grid. The two constitute the input and output core parameters of the energy conversion link of the wind farm. Real-time grid data covers key electrical quantities that characterize the operating status and frequency stability characteristics of the power system. The grid connection point system frequency refers to the AC frequency at the connection point between the wind farm and the grid, directly reflecting the active power supply and demand balance of the power system. The frequency change rate is the frequency change rate obtained by performing real-time differential calculations on the grid connection point system frequency, which is used to quantify the dynamic trend of the system frequency offset. The tie line power refers to the active power value that interacts between the region where the wind farm is located and the external grid through the tie line, reflecting the power transmission status between regions. The regional load refers to the total power load in the region where the wind farm is connected, which reflects the local active power demand. These parameters together constitute the key input variables for grid frequency stability analysis. In short, by obtaining real-time wind farm data and real-time grid data, a complete data chain covering the two-way information exchange between "new energy generation units and power systems" can be constructed. Specifically, real-time wind farm data is the basis for analyzing the fluctuation characteristics of wind power. By analyzing the temporal correlation between wind speed and power values, we can explore the dynamic conversion patterns between wind energy input and power output, providing the original driving signal for short-term power fluctuation prediction. Real-time grid data, on the other hand, is a key basis for assessing frequency stability requirements. System frequency and frequency change rate directly reflect the degree and dynamic trend of the current imbalance between active power supply and demand. Tieline power and regional load further reveal the spatial distribution of power imbalances and local demand characteristics. The integration of the two enables the regulation system to simultaneously perceive the "fluctuation characteristics of renewable energy generation" and the "frequency control requirements of the grid," avoiding the one-sided decision-making caused by reliance on a single data source. Without wind farm data, the potential impact of power fluctuations on frequency cannot be accurately predicted; without grid data, it is difficult to determine the actual demand for wind power regulation in the current system state. Through the collaborative analysis of the two types of data, the regulation system can establish a complete logical chain from "fluctuation prediction" to "impact assessment" to "strategy generation", realizing the transformation of wind power grid-connected control strategy from "wind turbine-centered local optimization" to "system-centered global coordination", and ultimately improving the response accuracy and system adaptability of wind farms in the frequency regulation process.

[0023] In this embodiment of the present application, the short-term power fluctuation prediction module 120 is configured to perform sequence encoding and decoding on the time series of the wind farm's real-time data to generate a prediction result for future short-term power fluctuations. It should be understood that the intermittent and fluctuating nature of wind energy results in instability and uncertainty in wind farm output power. This characteristic poses challenges to the stable operation of the power grid, particularly affecting the grid frequency. Only by accurately understanding the changing trends of wind power can appropriate measures be taken in advance to address power fluctuations and ensure safe and stable grid operation. By predicting short-term wind power fluctuations in advance, the system can adjust relevant control parameters, such as the virtual inertia coefficient and the reserved reserve percentage, before power changes. This allows wind turbines to participate in grid frequency regulation more promptly and accurately, reducing the impact of power fluctuations on the grid frequency and improving grid frequency stability. Traditional methods typically rely on simple statistical models or regression analysis based on historical data to estimate future power output. However, these methods often ignore the complex dynamic relationship between wind speed and power and their temporal trends, resulting in inaccurate prediction results, especially when dealing with severe fluctuations in wind power. Furthermore, traditional forecasting models lack a deep understanding of the real-time operating status of wind farms and are unable to effectively capture the interactions between the various factors that influence power output. This not only limits forecast accuracy but also makes it difficult for grid operators to take effective measures to mitigate the impact of wind power fluctuations on grid frequency stability.

[0024] Based on this, the technical concept of this application is to achieve high-precision short-term fluctuation forecasting through a time-series encoding and decoding architecture. First, the wind speed and power time series in real-time wind farm data are grouped separately. A recurrent neural network (RNN) is used to capture their respective temporal dependencies, generating a time-series correlation feature encoding vector for wind speed and power. Subsequently, fine-grained temporal collaborative reasoning is used to dynamically analyze the nonlinear coupling relationship between wind speed variation and power response, such as the dynamic correlation between gust acceleration and unit output lag. This constructs a fine-grained wind speed-power temporal response inference encoding vector. Finally, temporal decoding is performed based on this encoding vector to predict the range and rate of change in the wind farm's total active power in the future, which is extremely short-term. Specifically, in the technical solution of this application, the future short-term power fluctuation forecast includes the range and rate of change in the wind farm's total active power in the future, which is extremely short-term. This technical concept addresses the shortcomings of traditional models in capturing complex dynamic relationships by deeply exploring the dynamic interaction characteristics of wind speed and power and the multi-dimensional correlation between real-time operating conditions. For example, when wind speed fluctuates violently, the model can capture the impact of implicit factors such as unit inertial response and pitch control delay on power output through fine-grained reasoning, significantly improving the accuracy of short-term fluctuation prediction.

[0025] Specifically, Figure 2FIG. 1 is a block diagram of a short-term power fluctuation prediction module in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application. Figure 2 As shown, the short-term power fluctuation prediction module 120 includes: a wind farm real-time data grouping unit 121, which is used to group the time series of the wind farm real-time data according to the parameter sample dimension to obtain the time series of wind speed values ​​and the time series of power values; a wind farm real-time data encoding unit 122, which is used to perform sequence encoding on the time series of wind speed values ​​and the time series of power values ​​to obtain a wind speed value time series associated feature encoding vector and a power value time series associated feature encoding vector; a wind speed-power time series reasoning unit 123, which is used to perform time series fine-grained collaborative reasoning on the wind speed value time series associated feature encoding vector and the power value time series associated feature encoding vector to obtain a wind speed-power time series fine-grained response reasoning encoding vector; a power prediction unit 124, which is used to perform time series feature decoding and prediction on the wind speed-power time series fine-grained response reasoning encoding vector to obtain the future short-term power fluctuation prediction result.

[0026] In an embodiment of the present application, the wind farm real-time data grouping unit 121 is used to group the time series of the wind farm real-time data according to the parameter sample dimension to obtain the time series of wind speed values ​​and the time series of power values. Accordingly, it is considered that the accuracy of wind power prediction is highly dependent on the in-depth analysis of the dynamic correlation characteristics of wind speed and power. When processing multi-parameter data, traditional wind power prediction models often adopt a unified time series processing method to directly splice different physical quantities such as wind speed and power into the model. This extensive data processing method ignores the heterogeneous characteristics between different parameters - wind speed changes are dominated by atmospheric turbulence and present high-frequency randomness, while power response is affected by unit inertia and control strategy and has inertial delay characteristics. When the two are input into the model in the form of a mixed time series, it is difficult for the neural network to distinguish the intrinsic characteristics of the data source, resulting in the model confusing the causal relationship between the arrival time of the gust front and the pitch control delay during the learning process. Based on this, in order to extract features from the time series of wind speed values ​​and power values ​​respectively, the present application groups the time series of the real-time data of the wind farm according to the parameter sample dimension to decouple the wind speed and power from the original mixed data stream, forming independent time series analysis objects, and obtaining the time series of wind speed values ​​and the time series of power values.

[0027] In an embodiment of the present application, the wind farm real-time data encoding unit 122 is used to perform sequence encoding on the time series of the wind speed values ​​and the time series of the power values ​​to obtain a wind speed value time series correlation feature encoding vector and a power value time series correlation feature encoding vector. Specifically, in an embodiment of the present application, the wind farm real-time data encoding unit is used to perform sequence encoding on the time series of the wind speed values ​​and the time series of the power values ​​based on the RNN model to obtain a wind speed value time series correlation feature encoding vector and a power value time series correlation feature encoding vector. Accordingly, considering that the wind speed and power data of the wind farm are sequence data that change over time, they have strong time series correlation. For example, the wind speed and power at the current moment are closely related to the values ​​in the past period of time. If only statistical regression or a simple sliding average method is used, it is difficult for the model to parse such a causal chain across time steps, resulting in misjudgment of the power mutation trend. In addition, the wind speed sequence itself has non-stationary characteristics such as turbulent pulsation and gust clusters, while the power sequence is modulated by external factors such as the unit control logic and grid dispatch instructions. There are significant differences between the two in time scale and change mode, and their independent time series laws need to be modeled separately to extract effective features. To this end, in the technical solution of the present application, the time series of the wind speed value and the time series of the power value are sequence encoded based on the RNN model to obtain the wind speed value time series associated feature encoding vector and the power value time series associated feature encoding vector. Specifically, the recurrent feedback structure of the RNN (such as LSTM or GRU unit) is used to analyze the dynamic patterns such as turbulence intensity changes and gust duration in the wind speed sequence time step by time to generate a feature encoding vector that characterizes the evolution law of the wind speed time series; the power sequence is encoded synchronously to capture the influence of control dynamics such as the unit inertial response and active power regulation delay on the power output, and form a power time series associated feature vector. Through separate encoding, the model can independently learn the differentiated timing characteristics of wind speed disturbances and unit responses, such as identifying the phase difference between wind speed increase and power ramp in a gust event, providing a decoupled feature expression basis for subsequent interactive reasoning.

[0028] In the embodiment of the present application, the wind speed-power time series reasoning unit 123 is used to perform time series fine-grained collaborative reasoning on the wind speed value time series associated feature coding vector and the power value time series associated feature coding vector to obtain a wind speed-power time series fine-grained response reasoning coding vector. Specifically, Figure 3 FIG. 1 is a block diagram of a wind speed-power time series reasoning unit in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application. Figure 3As shown, the wind speed-power time series reasoning unit 123 includes: an implicit fine-grained association subunit 123-1, which is used to perform local time attention weighted implicit fine-grained association modulation on the wind speed value time series correlation feature coding vector and the power value time series correlation feature coding vector to obtain a set of wind speed-power local implicit time series correlation feature modulation coding vectors; a time series fine-grained response reasoning coding subunit 123-2, which is used to input the set of wind speed-power local implicit time series correlation feature modulation coding vectors into a time series reasoner based on a forward LSTM model to obtain the wind speed-power time series fine-grained response reasoning coding vector.

[0029] It should be understood that the relationship between wind speed and power is not a simple linear one, but rather a complex nonlinear coupling. For example, during a gust of wind, a wind turbine's power output does not immediately follow wind speed changes due to factors such as its own inertial response and pitch control delay. Instead, it experiences a certain lag and dynamic adjustment process. Traditional prediction models, lacking a deep understanding of the cross-temporal interaction between the two, struggle to accurately capture nonlinear effects such as power lag and inertial response during gust events. For example, when wind speed suddenly increases, the transfer of aerodynamic energy from the rotor to the generator through the drive train is subject to mechanical inertia delays, while the pitch control system requires seconds to tens of seconds to adjust the pitch angle to limit overspeed. These dynamic processes result in power output not being instantaneously synchronized with wind speed changes. Relying solely on simple wind speed-power mapping or static correlation models fails to capture these cross-timescale energy transfer and control delays. Especially during periods of severe wind speed fluctuations, models can easily misinterpret instantaneous wind speed peaks as sudden power changes, ignoring the buffering effect of the turbine's dynamic response, ultimately causing predictions to deviate from the actual power curve. Based on this, the present application performs time-series fine-grained collaborative reasoning on the wind speed value time-series correlation feature coding vector and the power value time-series correlation feature coding vector to accurately describe the dynamic relationship between wind speed and power, and obtains a wind speed-power time-series fine-grained response reasoning coding vector.

[0030] Specifically, Figure 4 FIG is a block diagram of an implicit fine-grained associated subunit in a wind power grid-connected regulation system based on multivariate data analysis according to an embodiment of the present application. Figure 4As shown, the implicit fine-grained association subunit 123-1 includes: a wind speed power local implicit feature extraction secondary subunit 123-11, which is used to perform one-dimensional convolution local implicit feature extraction on the wind speed value time series association feature coding vector and the power value time series association feature coding vector respectively to obtain a set of wind speed value local implicit time series association feature coding vectors and a set of power value local implicit time series association feature coding vectors; a wind speed-power time series feature interaction secondary subunit 123-12, which is used to perform single-point time series feature interaction on each corresponding group of wind speed value local implicit time series association feature coding vectors and power value local implicit time series association feature coding vectors in the set of wind speed value local implicit time series association feature coding vectors and the set of power value local implicit time series association feature coding vectors to obtain wind speed value local implicit time series association feature coding vectors. A set of wind speed-power local implicit temporal association feature coding vectors; a wind speed-power local temporal attention weight calculation secondary subunit 123-13, used to determine the local temporal attention weights of each wind speed-power local implicit temporal association feature coding vector in the set of wind speed-power local implicit temporal association feature coding vectors based on the characteristic distribution characteristics of each wind speed-power local implicit temporal association feature coding vector to obtain a set of wind speed-power local temporal attention weights; a wind speed-power local temporal modulation secondary subunit 123-14, used to weighted modulate the set of wind speed-power local implicit temporal association feature coding vectors based on the set of wind speed-power local temporal attention weights to obtain a set of wind speed-power local implicit temporal association feature modulation coding vectors.

[0031] More specifically, in the embodiment of the present application, the wind speed and power local implicit feature extraction secondary subunit 123-11 is used to perform one-dimensional convolution local implicit feature extraction on the wind speed value time series correlation feature coding vector and the power value time series correlation feature coding vector respectively to obtain a set of wind speed value local implicit time series correlation feature coding vectors and a set of power value local implicit time series correlation feature coding vectors. This process can be expressed by the formula:

[0032]

[0033]

[0034] in, is the temporal correlation feature encoding vector of wind speed value, is the power value temporal correlation feature encoding vector, It is one-dimensional convolution local implicit feature extraction, is the length of the one-dimensional convolution kernel, , , and They are the first, second, and third in the set of local implicit temporal correlation feature encoding vectors of wind speed values. and The local implicit temporal correlation feature encoding vector of wind speed values, , , and They are the first, second, and third in the set of local implicit temporal correlation feature encoding vectors of power values. and The power value local implicit temporal correlation feature encoding vector, yes and The number of vectors in and Same length.

[0035] It's understandable that time series data for wind speed and power contain complex local dynamic correlations, such as the delayed response of power during sudden wind gusts due to mechanical inertia and control delays. Traditional models, lacking the ability to analyze these local patterns, can lead to prediction bias. One-dimensional convolution techniques scan the time series correlation feature encoding vector in a parameter-sharing manner. Using convolution kernels of varying sizes (small kernels focus on fine-grained fluctuations, while large kernels integrate trends within a wider time window), they extract local implicit time series correlation features for both wind speed and power values, such as the slope of the upward trend during a sudden wind speed increase and the delayed onset of the power response, among other nonlinear dynamic features. This step converts the raw time series data into a set of feature encoding vectors embodying local interaction mechanisms, stripping away noise from the raw data to highlight cross-temporal coupling clues between wind speed variations and power responses (e.g., the delayed peaking of power after the wind speed peak). This provides a structured underlying feature representation for subsequent refined interaction modeling.

[0036] More specifically, in an embodiment of the present application, the wind speed-power time series feature interaction secondary subunit 123-12 is used to perform single-point time series feature interaction on each corresponding group of wind speed value local implicit time series association feature coding vectors and power value local implicit time series association feature coding vectors in the set of wind speed value local implicit time series association feature coding vectors and the set of power value local implicit time series association feature coding vectors to obtain a set of wind speed-power local implicit time series association feature coding vectors. This process can be expressed by the formula:

[0037]

[0038] in, It is the point product of position. It is added by position point. It is subtracted by position point, It is a cascade operation. is the set of local implicit temporal association weight matrices A local implicit temporal correlation weight matrix, is the set of local implicit temporal correlation bias vectors local implicit temporal correlation bias vector, is the first in the set of wind speed-power local implicit temporal correlation feature encoding vectors A wind speed-power local implicit temporal correlation feature encoding vector.

[0039] It should be understood that the nonlinear coupling of wind speed and power is reflected in the dynamic interaction at each local time point. For example, the restriction of the wind speed change rate at a certain moment and the power change at the corresponding moment requires modeling the point-by-point dependency between the two at the local implicit feature level. By performing a single feature interaction between the local implicit features of the wind speed value and the power value at each corresponding time point, the synergistic or inhibitory effect of the two on a fine-grained time scale can be captured (for example, when the wind speed rises rapidly, the power exhibits a nonlinear growth pattern due to the failure to adjust the pitch angle in time). In other words, the set of wind speed-power local implicit temporal correlation feature encoding vectors generated by this step can accurately characterize the dynamic modulation relationship of the wind speed feature on the power feature at each time point (such as the attenuation of wind speed energy transfer caused by mechanical transmission delay), avoid the noise redundancy problem caused by the direct splicing of the original features, and enable the interactive representation to focus on the energy conversion constraints of the wind speed-power cross-physical link, thereby providing a refined local interaction unit for subsequent temporal reasoning.

[0040] More specifically, in an embodiment of the present application, the wind speed-power local temporal attention weight calculation secondary subunit 123-13 is used to: determine the local temporal attention weight of each wind speed-power local implicit temporal association feature coding vector based on the characteristic distribution characteristics of each wind speed-power local implicit temporal association feature coding vector in the set of wind speed-power local implicit temporal association feature coding vectors to obtain a set of wind speed-power local temporal attention weights. This process can be expressed by the formula:

[0041]

[0042] in, is the first in the set of wind speed-power local implicit temporal correlation feature encoding vectors Wind speed-power local implicit temporal correlation feature encoding vector, yes Middle eigenvalues, To calculate the square of the Euclidean norm of a vector, yes The number of eigenvalues ​​in , yes function, is the first in the set of wind speed-power local temporal attention weights wind speed-power local temporal attention weights.

[0043] It should be understood that in the dynamic coupling process of wind speed and power, the contribution of local interaction features at different time points to the final power prediction varies (for example, the wind speed gradient change at the initial moment of a gust is more critical than the wind speed value in the steady-state stage). The importance of each local interaction feature needs to be evaluated based on the characteristic distribution characteristics. For example, when the wind speed change rate fluctuates abnormally, the interaction feature at the corresponding time point may indicate that the power is about to change dramatically and should be given a higher weight. In other words, by calculating the set of wind speed-power local temporal attention weights, the model can automatically identify the key time points that dominate power fluctuations (such as the features at the time when the pitch control action is triggered), suppress redundant information in steady-state or noise-dominated periods (such as the small fluctuation features at low wind speeds), and provide a selective attention mechanism for subsequent weighted modulation, ensuring that the reasoning process focuses on the local interaction features that carry the core dynamic information.

[0044] More specifically, in the embodiment of the present application, the wind speed-power local timing modulation secondary subunit 123-14 is used to: perform dynamic spatial constraint and period compensation on the set of wind speed-power local timing attention weights to obtain a set of wind speed-power local timing attention compensation weights. This process can be expressed as follows:

[0045] , ,

[0046]

[0047]

[0048]

[0049] in, is the first in the set of wind speed-power local time series interaction energy values The wind speed-power local time series interaction energy value, is the first in the set of wind speed-power local time series shift energy values. The wind speed-power local time series shift energy value, It is the first in the set of wind speed-power local time series oscillation energy values. The wind speed-power local time series oscillation energy value, The natural constant The logarithmic function value with base , It is the first factor in the set of wind speed-power local time series period rule compensation factors. A wind speed-power local timing cycle rule compensation factor, It is the first phase correction term in the set of wind speed-power local time series period. A wind speed-power local time series period phase correction term, and They are and The corresponding weight coefficient is is the first in the set of wind speed-power local temporal attention compensation weights Wind speed-power local temporal attention compensation weight;

[0050] Based on the set of wind speed-power local temporal attention compensation weights, the set of wind speed-power local implicit temporal correlation feature coding vectors is weighted modulated to obtain the set of wind speed-power local implicit temporal correlation feature modulation coding vectors. This process can be expressed as follows:

[0051]

[0052]

[0053] in, , , and They are the first, second, and third in the set of wind speed-power local implicit temporal correlation feature modulation coding vectors. and Wind speed-power local implicit temporal correlation feature modulation coding vector, It is a set of modulation coding vectors of the local implicit temporal correlation characteristics of wind speed and power.

[0054] It should be understood that weighted modulation of local implicit temporal correlation features based on attention compensation weights essentially enhances key information and suppresses noise. For example, when the wind speed-power interaction feature at a certain point in time corresponds to a critical stage of mechanical inertia delay (such as a sudden torque change in the drive train), its high attention weight amplifies the feature's impact on subsequent reasoning, while low-weight features (such as sensor measurement noise) are weakened. In other words, the resulting set of modulated encoding vectors of local implicit temporal correlation features of wind speed and power can highlight core dynamic elements of the wind speed-power coupling process (such as the starting point of power response lag during a gust of wind and the inflection point of pitch control). Ultimately, this forms a structured set of temporal features. Each modulated encoding vector of a local implicit temporal correlation feature of wind speed and power contains both local interaction information and its importance weight in the global temporal sequence, providing hierarchical input data for modeling long-range dependencies in the LSTM model.

[0055] In particular, when calculating each wind speed-power local implicit time series correlation feature coding vector, the feature interaction between the corresponding wind speed value local implicit time series correlation feature coding vector and the power value local implicit time series correlation feature coding vector is introduced. For example, , , Etc., which actually correspond to different interaction paradigms, thus forming differentiated spatial constraint association mechanisms in the interaction space.

[0056] In order to enhance the paradigm dynamic adaptability of the wind speed-power local temporal attention weight, it is preferred to modify the weight based on the spatial effect analysis of the interaction paradigm. Specifically, , is considered as a translation effect (i.e., the gradient direction is consistent with the spatial interaction direction), and is considered as an oscillatory effect (i.e. the gradient direction is orthogonal to the spatial interaction direction). Vector statistics for different types of interactions (e.g. , , ), the oscillation effect induces a local periodic response phenomenon in the translation direction, and the compensation factor of the wind speed-power local time series periodic rule needs to be calculated:

[0057]

[0058] in, As a translation domain representation, the intensity of the oscillation effect It increases with the growth of the logarithmic dimension of the translation domain. At the same time, the oscillation effect will also cause the phase jump of the translation domain, thereby generating the phase correction term of the local time series period of wind speed-power:

[0059] Finally, the wind speed-power local time series attention weight is modified based on the weighted sum of the above two items :

[0060] This method strengthens the correlation of the spatial interaction rule system by distinguishing the different roles of interaction paradigms in the spatial constraint mechanism, thereby improving the calculation accuracy of the wind speed-power local temporal attention weight.

[0061] Specifically, in the embodiment of the present application, the temporal fine-grained response reasoning encoding subunit 123-2 is used to input the set of the wind speed-power local implicit temporal correlation feature modulation coding vectors into a temporal reasoner based on a forward LSTM model to obtain the wind speed-power temporal fine-grained response reasoning coding vector. This process can be expressed as follows:

[0062]

[0063] in, is the forward LSTM encoding, is the wind speed-power time series fine-grained response inference encoding vector.

[0064] It's understandable that the nonlinear coupling of wind speed and power across timescales (e.g., from second-level wind speed changes to tens-of-second power adjustments) requires a model to capture long-range temporal dependencies. The LSTM's gating mechanism (forget gate, input gate, and output gate) can effectively handle this type of non-stationary sequence data, memorizing long-term influencing factors such as mechanical inertia delay and pitch control dynamics. Furthermore, its forward structure ensures that the reasoning process proceeds chronologically, conforming to the causal logic of physical processes (wind speed changes first, followed by power response). Specifically, through the LSTM's chained reasoning, dispersed local modulation features are integrated into a global, fine-grained inference encoding vector for the wind speed-power time series response. This vector not only contains fine-grained interaction information at each time point (e.g., the nonlinear mapping between the wind speed rate of change and the power rate of change at a given moment), but also captures dynamic evolution patterns across time periods (e.g., how power is constrained by the turbine control strategy as wind speed continues to rise, gradually approaching its limit). Compared with the instantaneous synchronization assumption of traditional models, this step can accurately characterize nonlinear effects such as lag and saturation between wind speed and power, and provide a deep feature representation that includes physical process constraints for subsequent power fluctuation prediction.

[0065] In an embodiment of the present application, the power prediction unit 124 is used to perform time series feature decoding and prediction on the wind speed-power time series fine-grained response reasoning coding vector to obtain the future short-term power fluctuation prediction result. Specifically, in an embodiment of the present application, the power prediction unit is used to: perform RNN-based time series feature decoding and prediction on the wind speed-power time series fine-grained response reasoning coding vector to obtain the future short-term power fluctuation prediction result, and the future short-term power fluctuation prediction result includes the future very short-term wind farm total active power change range and rate. It should be understood that the wind speed-power time series fine-grained response reasoning coding vector contains the complex dynamic relationship between wind speed and power and the characteristic information on the time series. The RNN-based time series feature decoding and prediction can further explore these potential laws and convert the information in the coding vector into understandable and applicable future short-term power fluctuation prediction results. The RNN model has the ability to process sequence data and can predict future states based on historical information and the current coding state. It is suitable for decoding such coding vectors with time series characteristics. This prediction of future short-term power fluctuations can provide a key basis for decision-making in wind power grid-connected regulation systems. Specifically, it clearly defines the range and rate of change in the wind farm's total active power over very short periods of time. This helps grid operators and wind power companies understand the wind farm's power output in advance, allowing them to take appropriate measures to address power fluctuations.

[0066] In particular, in a specific embodiment of the present application, the implementation process of performing RNN-based time series feature decoding prediction on the wind speed-power time series fine-grained response inference encoding vector to obtain a future short-term power fluctuation prediction result including the change range and rate of the future extremely short-term wind farm total active power includes:

[0067] First, the model architecture must be selected and the parameters configured. Given the nonlinear and non-stationary nature of wind power time series data, it is necessary to select an RNN variant architecture suitable for processing long sequence dependencies, such as LSTM or GRU. If LSTM is selected, its core parameters must be determined. The number of hidden layers is generally set to 1-3 layers, which is to balance the complexity and generalization ability of the model. The number of neurons in each layer can be set initially based on the length of historical data and feature dimensions through experience or preliminary experiments. The learning rate can adopt an adaptive strategy, and the number of iterations is set to 50-200 training cycles. In order to obtain the optimal parameter combination, cross-validation is required. The historical data is divided into training, validation, and test sets. Using grid search or random search methods, the parameters are determined with the goal of minimizing the loss of the validation set.

[0068] Next, we preprocess the input data. The model uses the encoded vector of the wind speed-power time series fine-grained response inference as input. Because the dimensions of different features may vary, the input data needs to be normalized to scale it to an appropriate range. Normalization accelerates model convergence and prevents training instability caused by large feature value differences.

[0069] During the model initialization and time series processing phases, the LSTM's hidden state and cell state must be initialized before training. Both are zero vectors with dimensions equal to the number of neurons in the hidden layer. The encoded vector data for each time step is then fed into the model sequentially, following the time series order. At each time step, the LSTM performs calculations using the input gate, forget gate, and output gate. The input gate determines how much information from the current input data is stored in the cell state; the forget gate controls how much information from the previous cell state is forgotten; and the output gate determines the output content of the current hidden state. Through these gating operations, the model can capture long-term dependencies in time series data.

[0070] After the LSTM processes all time-step input data, it extracts the hidden state at the last time step and feeds it into the fully connected layer. The fully connected layer uses a linear transformation to map the hidden state to the prediction space, generating a preliminary prediction. Because the input data is normalized, the prediction results must be denormalized to restore their true physical dimensions.

[0071] Regarding loss function definition and parameter optimization, a loss function is defined to measure the error between the predicted value and the true value. Mean squared error (MSE) or mean absolute error (MAE) is commonly used. An adaptive optimization algorithm is used to update model parameters. The loss error is propagated backward through the network via backpropagation, adjusting the parameters of each layer to minimize the loss function. During training, model performance is regularly evaluated on the validation set. If the validation set loss does not decrease for several consecutive epochs, training is terminated early to avoid overfitting.

[0072] Finally, the model is validated and results are output. After training is complete, the model is evaluated using a test set to verify its generalization ability on unseen data. Evaluation metrics include the deviation between the predicted value and the true value and the degree of consistency with the trend. If performance meets the requirements, the model is put into practical use. During the real-time operation phase, the current wind speed-power time series fine-grained response inference encoding vector is input. After the decoding process described above, the range and rate of change of the wind farm's total active power in the future in the short term are output, providing an accurate prediction basis for subsequent frequency impact assessment and control strategy generation.

[0073] In summary, the short-term power fluctuation prediction module has been clearly explained. It first groups the wind speed and power time series in the real-time data of the wind farm, then uses a recurrent neural network to capture their respective temporal dependencies to generate a temporal correlation feature encoding vector for wind speed and power. Subsequently, through fine-grained temporal collaborative reasoning, it dynamically analyzes the nonlinear coupling relationship between wind speed changes and power responses, thereby constructing a wind speed-power temporal fine-grained response reasoning encoding vector. Finally, based on this encoding vector, temporal decoding is performed to predict the range and rate of change of the total active power of the wind farm in the future in the extremely short term. In this way, by deeply exploring the dynamic interactive characteristics of wind speed and power and the multi-dimensional correlation of real-time operating status, the shortcomings of traditional models in insufficiently characterizing complex dynamic relationships can be effectively addressed, thereby significantly improving the accuracy of short-term fluctuation prediction.

[0074] In an embodiment of the present application, the expected frequency impact analysis module 130 is used to input the future short-term power fluctuation prediction result and the real-time data of the power grid into a frequency response analysis model to obtain an expected frequency impact evaluation result. Specifically, in an embodiment of the present application, the expected frequency impact analysis module is used to: perform LSTM-based time series encoding on the future short-term power fluctuation prediction result to obtain a future short-term power fluctuation time series encoding vector; perform MLP-based data encoding on the real-time data of the power grid to obtain a real-time data encoding vector of the power grid; and fuse the future short-term power fluctuation time series encoding vector and the real-time data encoding vector of the power grid and input them into a decoder-based frequency evaluator to obtain the expected frequency impact evaluation result. Here, the frequency response analysis model includes LSTM, MLP, and a decoder-based frequency evaluator.

[0075] As you can understand, the future short-term power fluctuation forecast results, as time series data, contain key information such as the range and rate of change of the wind farm's total active power in the very short term. The purpose of using LSTM (Long Short-Term Memory) networks for time series encoding is to deeply explore the temporal dependencies and changing trends in this data. Through its unique gating mechanism, LSTMs can effectively process long-term data, selectively remembering or forgetting information, thereby better capturing both long-term and short-term characteristics of power fluctuations. After LSTM processing, the original complex data is converted into a low-dimensional vector representation, namely the time series encoding vector of the future short-term power fluctuations.

[0076] In particular, real-time power grid data encompasses a wide range of information reflecting the current operating status of the power grid, including system frequency at the grid connection point, frequency change rate, tie-line power, and regional load. To transform this high-dimensional, complex real-time power grid data into a more manageable form, an MLP (Multi-Layer Perceptron) is employed for data encoding. As a feedforward neural network, the MLP automatically learns feature representations in the data through nonlinear transformations in multiple hidden layers, encoding the real-time power grid data into a compact and representative vector—the real-time power grid data encoding vector. This approach not only reduces data dimensionality and computational complexity but also extracts features critical for frequency assessment, enabling subsequent frequency assessment models to process this information more efficiently. Therefore, the MLP data encoding process effectively extracts and compresses the real-time power grid data, converting key frequency-related features into a vector form suitable for model processing. This improves the accuracy and efficiency of frequency assessment while avoiding overfitting and computational inefficiencies caused by overly complex raw data.

[0077] The time-series code vectors of future short-term power fluctuations are then fused with the code vectors of real-time grid data to comprehensively consider the combined impact of the wind farm's future power fluctuations and the current grid operating state on the grid frequency. A decoder-based frequency estimator, building on this foundation, assesses the expected impact of future short-term power fluctuations on the grid frequency based on the fused code vectors. The decoder is a neural network structure capable of mapping code vectors back to the original data space or performing task-specific predictions. In this scenario, it is used to predict the expected frequency impact assessment results, such as the grid frequency variation range and the likelihood of frequency fluctuation, based on the fused feature vectors. By combining the two code vectors, the model fully leverages the time-series characteristics of wind farm power fluctuations and the characteristics of the real-time grid operating state to accurately assess the impact of future short-term power fluctuations on the grid frequency. Specifically, the model analyzes the frequency deviation and frequency variation rate that may be caused by the predicted power variation (i.e., the future short-term power fluctuation prediction result package) under the current grid operating state (i.e., the grid connection point system frequency, frequency variation rate, tie line power, and regional load in the real-time grid data).

[0078] In this embodiment of the present application, the level generation module 140 is configured to determine a frequency risk level based on the future short-term power fluctuation prediction results and the expected frequency impact assessment results. It is understood that the future short-term power fluctuation prediction results provide information such as the range and rate of change in the wind farm's total active power over a very short period of time in the future, reflecting the volatility of the wind farm's own power output, a key factor affecting grid frequency stability. Because a wind farm is part of the power grid, rapid changes in its power may have a direct impact on the grid frequency. The expected frequency impact assessment results comprehensively consider future short-term power fluctuations and the real-time operating status of the grid, assessing the potential impact on the grid frequency. They more comprehensively reflect the grid frequency changes that may be caused by wind farm power fluctuations under current grid conditions, including the frequency range and the likelihood of frequency fluctuations, and serve as a direct basis for determining the frequency risk level. By combining the future short-term power fluctuation prediction results with the expected frequency impact assessment results and applying pre-set assessment criteria and thresholds, the frequency risk level can be determined. This grading approach helps quickly understand the degree of risk facing the grid frequency, enabling timely implementation of appropriate measures to ensure the safe and stable operation of the grid.

[0079] In particular, in a specific embodiment of the present application, the future short-term power fluctuation prediction results and the expected frequency impact assessment results can be input into a rule base to obtain the frequency risk level, wherein the frequency risk level includes low risk, medium risk, high risk and extremely high risk.

[0080] Specifically, if the minimum frequency predicted in the expected frequency impact assessment result is higher than the lower limit of normal operation, and the absolute value of the frequency change rate predicted in the expected frequency impact assessment result is lower than a certain safety threshold, it is judged to be a low risk. If the minimum frequency predicted in the expected frequency impact assessment result may fall to near or slightly below the lower limit of normal operation, or the frequency change rate predicted in the expected frequency impact assessment result is close to the safety threshold, it is judged to be a medium risk. If the frequency predicted in the expected frequency impact assessment result may fall below the lower limit of allowable deviation, or the frequency change rate predicted in the expected frequency impact assessment result significantly exceeds the safety threshold, it is judged to be a high risk. If the frequency predicted by the minimum frequency predicted in the expected frequency impact assessment result may fall below the emergency measures threshold, it is judged to be an extremely high risk.

[0081] It's worth noting that in this specific example, the comparison and mapping not only considers the predicted frequency extremes and rates of change, but also the predicted power fluctuation magnitude and rate itself. For example, even if the predicted frequency impact is moderate, an unusually fast rate of power change could raise the risk level, as rapid changes inherently increase uncertainty and control difficulty.

[0082] In another specific implementation of this application, a machine learning classification model can be used as a frequency risk assessment model. Its inputs are the predicted results of future short-term power fluctuations and the expected frequency impact assessment results, and its output is the frequency risk level. In this specific example, building a complete and accurate frequency risk assessment model is fundamental. This requires the collection of a large amount of rich historical data. This data covers wind power fluctuations and corresponding grid frequency changes in various scenarios, as well as grid operating status information at different times, such as the system frequency at the grid connection point, the frequency change rate, the tie line power, and the regional load. It is also necessary to fully consider the unique characteristics of wind power itself, such as its inherent intermittency and volatility, as well as the existing frequency regulation capabilities of the grid. By deeply analyzing this massive amount of historical data and combining it with multiple simulation experiments, the parameters and weights in the model can be clarified. For example, factors that have historically frequently caused large frequency fluctuations are assigned higher weights in the model to highlight their significant impact on frequency risk; factors with relatively small frequency impacts are assigned lower weights to ensure that the model accurately reflects the actual role of different factors in frequency risk assessment.

[0083] Next, the predicted results of future short-term power fluctuations are input into the established assessment model. This prediction contains a lot of key information, such as the predicted power change range, change rate, and uncertainty of power fluctuations. The power change range and change rate are crucial to the potential impact on grid frequency. When it is predicted that wind power will rise or fall significantly in the short term, and the speed of change exceeds the grid's normal regulation capabilities, this situation is likely to have a significant impact on the grid frequency. This is because the grid may find it difficult to adjust quickly when faced with such rapid and large power changes, which may cause the frequency to deviate from the normal operating range and increase frequency risk.

[0084] The model then incorporates the results of the expected frequency impact assessment. This assessment combines power fluctuation forecasts with real-time grid operating data, including key information such as the likely frequency range, the probability of the frequency exceeding the normal operating range, and the severity of the frequency fluctuations. For example, if the assessment indicates a high probability of the frequency falling below the minimum allowable value, or if the frequency fluctuations are likely to exceed the tolerance of grid equipment, these clearly indicate a high risk to the grid frequency, requiring significant attention and response.

[0085] The assessment model comprehensively analyzes various indicators from the forecast results of future short-term power fluctuations and the expected frequency impact assessment results. A weighted average approach can be used to assign weights to each indicator based on its importance to frequency risk, thereby calculating a comprehensive risk index. For example, the probability of frequency exceeding the normal range, a key indicator directly related to the safe operation of the power grid, is given a higher weight, while less important indicators are given lower weights. Furthermore, machine learning algorithms can be used, leveraging their powerful data processing and pattern recognition capabilities to enable the model to learn risk patterns from large amounts of historical data, thereby automatically and accurately identifying and determining the current frequency risk level. Machine learning algorithms can effectively handle complex nonlinear relationships and, when faced with numerous complex factors, can more accurately assess frequency risk than traditional methods.

[0086] Finally, the calculated comprehensive risk index is compared with pre-set risk level thresholds. These thresholds are determined based on multiple factors, including grid safety operation standards, equipment characteristics, and relevant industry regulations. Frequency risk levels are typically categorized as low, medium, high, and severe. When the comprehensive risk index falls below the low-risk threshold, the grid frequency is relatively stable, requiring standard monitoring and maintenance procedures. If the index falls within the medium-risk threshold, the risk is considered medium, necessitating enhanced monitoring of the grid frequency and the preparation of appropriate frequency regulation measures. If the index exceeds the high-risk threshold, the frequency risk is high, requiring immediate activation of emergency response plans and swift implementation of effective frequency regulation measures to ensure safe and stable grid operation. If the index reaches the severe-risk threshold, the situation becomes critical, requiring more urgent and robust measures to restore grid frequency stability and avoid serious power outages.

[0087] In an embodiment of the present application, the control instruction parameter generation module 150 is used to input the future short-term power fluctuation prediction results, the expected frequency impact assessment results, the frequency risk level, and the real-time grid data into the active frequency support strategy generation module to obtain control instruction parameters. The control instruction parameters include a virtual inertia coefficient, a reserved reserve percentage, a power change rate limit value, and a target active power adjustment value. The intermittent and uncertain nature of wind power can cause fluctuations in grid frequency. By inputting the above data into the active frequency support strategy generation module, appropriate control instruction parameters, such as the virtual inertia coefficient, the reserved reserve percentage, the power change rate limit value, and the target active power adjustment value, are generated. This can adjust the operation of the wind turbine to provide corresponding support when the frequency changes, simulate the inertial response of the synchronous generator, release the reserve power, or smooth the power change curve, thereby effectively suppressing frequency fluctuations and maintaining the grid frequency within a stable range.

[0088] In particular, in a specific embodiment of the present application, the specific implementation process of inputting the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level, and the real-time power grid data into the active frequency support strategy generation module to obtain the control instruction parameters includes:

[0089] The active frequency support strategy generation module first integrates and analyzes all input information. The frequency risk level allows the module to determine the current level of frequency stability pressure facing the power grid. The power change amplitude and rate in the future short-term power fluctuation forecast results can reflect the dynamic changes in wind power. The possibility and range of frequency offsets in the expected frequency impact assessment results allow the module to predict possible abnormalities in the grid frequency. The real-time grid data, which shows the operating status of the grid connection point system frequency, frequency change rate, tie line power, and regional load, provides real-time and accurate basic information for the calculation of subsequent parameters.

[0090] When calculating the virtual inertia coefficient, the module considers the grid frequency change rate and trend, combined with available regulation capacity. If the grid frequency change rate is high and trending downward, indicating challenges with grid frequency stability, the module increases the virtual inertia coefficient, enabling the wind turbine to enhance its ability to simulate the inertial response of a synchronous generator, quickly providing power during the initial stages of frequency fluctuations to stabilize the grid frequency. If the frequency change rate is low, indicating relatively stable grid frequency, the module appropriately decreases the virtual inertia coefficient to avoid overregulation.

[0091] In a specific example, the process of inputting the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid into the active frequency support strategy generation module to obtain the virtual inertia coefficient includes: obtaining a basic virtual inertia coefficient from a query table based on the frequency risk level; adjusting the basic virtual inertia coefficient according to the expected frequency conversion rate in the expected frequency impact assessment result to obtain a first adjustment item; calculating a second adjustment item based on the real-time power grid system inertia in the real-time power grid data; and adding the basic virtual inertia coefficient, the first adjustment item and the second adjustment item to obtain the virtual inertia coefficient.

[0092] The reserve reserve percentage is calculated primarily based on the frequency risk level and forecasts of future power fluctuations. When the risk level is high, the module increases the reserve reserve percentage to mitigate potential significant frequency fluctuations, allowing the wind turbine to operate at a slightly lower load, preserving more active power. Furthermore, a reasonable trigger threshold is set based on the expected frequency impact assessment results. When the grid frequency falls below this threshold, the wind turbine can quickly release its reserved reserve power.

[0093] In a specific implementation, the process of inputting the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid into the active frequency support strategy generation module to obtain the reserved reserve percentage includes: obtaining the basic standby power demand from the query table based on the frequency risk level; calculating the adjustment item based on the predicted frequency deviation based on the predicted frequency deviation in the expected frequency impact assessment result; calculating the adjustment item based on the predicted wind power based on the variation range of the total active power of the wind farm in the future short-term power fluctuation prediction result; adding the basic standby power demand, the adjustment item based on the predicted frequency deviation and the adjustment item based on the predicted wind power to obtain the total required standby power; dividing the total required standby power by the current maximum available output to obtain the reserved standby percentage.

[0094] The power rate of change limit is determined based on the severity of power fluctuations predicted in future short-term power fluctuation forecasts, as well as the line carrying capacity and equipment tolerances reported in real-time grid data. If a dramatic power fluctuation is predicted, the module calculates a lower power rate of change limit, proactively smoothing the power variation curve to prevent sudden power shocks to the grid. If power fluctuations are relatively stable, the module appropriately relaxes the power rate of change limit to maximize the wind turbine's generating capacity.

[0095] In a specific implementation, the process of inputting the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid into the active frequency support strategy generation module to obtain the power change rate limit value includes: based on the frequency risk level, obtaining the basic descent rate limit value from the query table; based on the predicted frequency change rate in the expected frequency impact assessment result, calculating the adjustment item based on the predicted frequency change rate; based on the change rate of the total active power of the wind farm in the future short-term power fluctuation prediction result, calculating the adjustment item based on the predicted wind power change rate; adding the basic descent rate limit value, the adjustment item based on the predicted frequency change rate and the adjustment item based on the predicted wind power change rate to obtain the power change rate limit value.

[0096] Calculating the target active power adjustment value is complex and requires considering multiple factors. First, the direction and extent of the impact of wind farm power output on grid frequency are determined by referring to forecasts of future short-term power fluctuations. If the predicted power increase causes the frequency to overshoot, the target active power value is reduced; otherwise, it is increased. Secondly, based on the expected frequency impact assessment results, if the assessment indicates a significant risk of the grid frequency deviating from the target value, the target active power adjustment range is increased accordingly. Finally, based on real-time grid data, including the system frequency at the connection point, frequency change rate, tie line power, and regional load, the target active power adjustment value is fine-tuned while ensuring safe and stable grid operation. For example, if tie line power approaches its transmission limit and regional load continues to increase, the target active power value may need to be significantly reduced if wind farm power fluctuations could increase the grid burden. Conversely, when grid power is sufficient, the target active power value can be appropriately increased to fully utilize wind energy resources.

[0097] In a specific implementation, the process of inputting the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid into the active frequency support strategy generation module to obtain the target active power adjustment value includes: obtaining a basic adjustment gain from a query table based on the frequency risk level; calculating an adjustment item based on the predicted frequency deviation based on the predicted frequency deviation in the expected frequency impact assessment result; calculating an adjustment item based on the predicted wind power change rate based on the variation range of the total active power of the wind farm in the future short-term power fluctuation prediction result; adding the adjustment item based on the predicted frequency deviation to the adjustment item based on the predicted wind power change rate, and then multiplying the sum by the basic adjustment gain to obtain the target active power adjustment value.

[0098] It's worth noting that in the aforementioned examples, the active frequency support strategy generation module is essentially a regression model based on a statistical model. Its inputs are the future short-term power fluctuation forecast, the expected frequency impact assessment, the frequency risk level, and real-time grid data. Its outputs are the virtual inertia coefficient, the reserved standby percentage, the power rate of change limit, and the target active power adjustment value. The calculation principles of each statistical model are presented in the specific examples.

[0099] It is also worth mentioning that in another embodiment of the present application, the active frequency support strategy generation module can also be a regression prediction model based on a machine learning model. For example, in a specific example, the active frequency support strategy generation module includes an encoder and a decoder, wherein the encoder is used to perform structured encoding on the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid to obtain the future short-term power fluctuation prediction result structured encoding vector, the expected frequency impact assessment result structured encoding vector, the frequency risk level structured encoding vector and the real-time data structured encoding vector of the power grid; and the decoder is used to perform vector regression decoding on the concatenated vector of the future short-term power fluctuation prediction result structured encoding vector, the expected frequency impact assessment result structured encoding vector, the frequency risk level structured encoding vector and the real-time data structured encoding vector of the power grid to obtain the virtual inertia coefficient, the reserved standby percentage, the power change rate limit value and the target active power adjustment value. In this specific example, the active frequency support strategy generation module is an end-to-end regression model, and its model parameters are obtained through a training process.

[0100] In this embodiment of the present application, the wind turbine adjustment module 160 is configured to transmit the control command parameters to the wind farm coordination controller, which then adjusts the physical behavior of the wind turbines based on the control command parameters. Specifically, by adjusting the physical behavior of the wind turbines, such as changing the virtual inertia coefficient, controlling the release of reserved standby power, and limiting the power change rate, the module can rapidly provide or absorb power when the grid frequency fluctuates, simulating the characteristics of synchronous generators, effectively supporting the grid frequency, and maintaining it within a stable range, thereby optimizing the overall performance of wind power grid integration. Specifically, the physical behavior adjustments of the wind turbines include virtual inertia adjustment, standby power release, power change rate limitation, and target power adjustment. Virtual inertia adjustment involves dynamically increasing or decreasing the output power of the wind turbines through the power electronic converter based on the virtual inertia coefficient in the control command parameters, thereby achieving a rapid response in the early stages of frequency fluctuations and simulating the inertia effect. For example, when the system frequency decreases, the coordination controller issues a command to the wind turbines to increase their output power; when the frequency increases, the controller issues a command to the wind turbines to decrease their output power. Reserve power release refers to the triggering of the wind turbine's coordinated controller to release reserved reserve active power when the grid frequency fluctuates significantly (e.g., when the frequency drops rapidly), rapidly injecting additional power into the grid to suppress further frequency drops. Power rate of change limitation refers to limiting the rate of change of the wind turbine's output power based on the power rate of change limit value in the control instruction parameters to ensure a smooth power regulation process and avoid impacts on the grid. Target power adjustment refers to adjusting the wind turbine's output power based on the target active power adjustment value in the control instruction parameters to achieve a specified target output power value.

[0101] In summary, the wind power grid-connected regulation system 100 based on multivariate data analysis according to the embodiment of the present application is explained, which first obtains the real-time data of the wind farm and the power grid, then uses the time series encoding and decoding technology to predict the future short-term power fluctuations of the wind farm, and combines the real-time data of the power grid to perform the expected frequency impact assessment, and then determines the frequency risk level based on the power fluctuation prediction and frequency impact assessment results, and then inputs the prediction results, assessment results, risk level and real-time data of the power grid into the active frequency support strategy generation module to generate control instruction parameters, and finally sends the control instruction parameters to the wind farm coordination controller for adjusting the physical action of the wind turbine to achieve active frequency support. In this way, it is conducive to improving the efficiency of wind power grid connection in participating in frequency regulation, and can effectively provide technical guarantees for the stable operation of a high proportion of new energy power grids.

Claims

1. A wind power grid-connected regulation system based on multivariate data analysis, characterized in that: include: A multivariate data acquisition module, configured to acquire multivariate data, wherein the multivariate data includes real-time data of wind farms and real-time data of power grids; A short-term power fluctuation prediction module is used to perform sequence encoding and decoding on the time series of the real-time data of the wind farm to obtain a prediction result of future short-term power fluctuations; an expected frequency impact analysis module, configured to input the future short-term power fluctuation prediction result and the real-time power grid data into a frequency response analysis model to obtain an expected frequency impact assessment result; a level generating module, configured to determine a frequency risk level based on the future short-term power fluctuation prediction result and the expected frequency impact assessment result; a control instruction parameter generation module, configured to input the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level, and the real-time grid data into an active frequency support strategy generation module to obtain control instruction parameters, wherein the control instruction parameters include a virtual inertia coefficient, a reserved standby percentage, a power change rate limit value, and a target active power adjustment value; The wind turbine adjustment module is configured to send the control instruction parameters to a wind farm coordination controller, and the wind farm coordination controller adjusts the physical actions of the wind turbine based on the control instruction parameters.

2. The wind power grid-connected regulation system based on multivariate data analysis according to claim 1, characterized in that: The real-time data of the wind farm includes wind speed value and power value; the real-time data of the power grid includes the system frequency of the grid connection point, the frequency change rate, the tie line power and the regional load.

3. The wind power grid-connected regulation system based on multivariate data analysis according to claim 2, characterized in that: The short-term power fluctuation prediction module includes: A wind farm real-time data grouping unit, configured to group the time series of the wind farm real-time data according to the parameter sample dimension to obtain a time series of wind speed values ​​and a time series of power values; A wind farm real-time data encoding unit, configured to perform sequence encoding on the time series of the wind speed value and the time series of the power value to obtain a wind speed value time series correlation feature encoding vector and a power value time series correlation feature encoding vector; A wind speed-power time series reasoning unit, configured to perform time series fine-grained collaborative reasoning on the wind speed value time series associated feature coding vector and the power value time series associated feature coding vector to obtain a wind speed-power time series fine-grained response reasoning coding vector; A power prediction unit, configured to perform time series feature decoding and prediction on the wind speed-power time series fine-grained response inference coding vector to obtain the future short-term power fluctuation prediction result; The wind farm real-time data encoding unit is configured to: perform sequence encoding based on an RNN model on the time series of the wind speed values ​​and the time series of the power values ​​to obtain a time series associated characteristic encoding vector of the wind speed values ​​and a time series associated characteristic encoding vector of the power values, wherein the RNN's cyclic feedback structure is utilized to analyze dynamic patterns such as turbulence intensity changes and gust duration in the wind speed sequence time step by time to generate a characteristic encoding vector that characterizes the evolution law of the wind speed time series; and simultaneously encode the power sequence to capture the influence of control dynamics such as unit inertial response and active power regulation delay on power output to form a power time series associated characteristic vector; Among them, the wind speed-power time series reasoning unit includes: an implicit fine-grained association subunit, which is used to perform local time attention weighted implicit fine-grained association modulation on the wind speed value time series association feature coding vector and the power value time series association feature coding vector to obtain a set of wind speed-power local implicit time series association feature modulation coding vectors; a time series fine-grained response reasoning coding subunit, which is used to input the set of wind speed-power local implicit time series association feature modulation coding vectors into a time series reasoner based on a forward LSTM model to obtain the wind speed-power time series fine-grained response reasoning coding vector, wherein, through the chain reasoning of LSTM, the scattered local modulation features are integrated into a global wind speed-power time series fine-grained response reasoning coding vector, which not only contains the fine-grained interaction information of each time point, but also captures the dynamic evolution law across time periods, wherein the fine-grained interaction information includes: a nonlinear mapping of the wind speed change rate and the power change rate.

4. The wind power grid-connected regulation system based on multivariate data analysis according to claim 3, characterized in that: The implicit fine-grained association subunit includes: a secondary subunit for extracting local implicit features of wind speed and power, configured to perform one-dimensional convolution local implicit feature extraction on the wind speed value temporal correlation feature coding vector and the power value temporal correlation feature coding vector, respectively, to obtain a set of wind speed value local implicit temporal correlation feature coding vectors and a set of power value local implicit temporal correlation feature coding vectors; a wind speed-power time series feature interaction secondary subunit, configured to perform single-point time series feature interaction on each corresponding group of wind speed value local implicit time series association feature coding vectors and power value local implicit time series association feature coding vectors in the set of wind speed value local implicit time series association feature coding vectors and the set of power value local implicit time series association feature coding vectors, respectively, to obtain a set of wind speed-power local implicit time series association feature coding vectors; a wind speed-power local temporal attention weight calculation secondary subunit, configured to determine, based on the characteristic distribution characteristics of each wind speed-power local implicit temporal association feature coding vector in the set of wind speed-power local implicit temporal association feature coding vectors, the local temporal attention weight of each wind speed-power local implicit temporal association feature coding vector to obtain a set of wind speed-power local temporal attention weights; The wind speed-power local timing modulation secondary subunit is used to weightedly modulate the set of wind speed-power local implicit timing association feature coding vectors based on the set of wind speed-power local timing attention weights to obtain the set of wind speed-power local implicit timing association feature modulation coding vectors.

5. The wind power grid-connected regulation system based on multivariate data analysis according to claim 4 is characterized in that: The wind speed-power local timing modulation secondary subunit is used to: Performing dynamic spatial constraint and period compensation on the set of wind speed-power local temporal attention weights to obtain a set of wind speed-power local temporal attention compensation weights; Based on the set of wind speed-power local temporal attention compensation weights, the set of wind speed-power local implicit temporal correlation feature coding vectors is weighted modulated to obtain the set of wind speed-power local implicit temporal correlation feature modulation coding vectors.

6. The wind power grid-connected regulation system based on multivariate data analysis according to claim 5, characterized in that: The power prediction unit is used to perform RNN-based time series feature decoding and prediction on the wind speed-power time series fine-grained response inference coding vector to obtain the future short-term power fluctuation prediction result, and the future short-term power fluctuation prediction result includes the future extremely short-term wind farm total active power change range and rate.

7. The wind power grid-connected regulation system based on multivariate data analysis according to claim 6, characterized in that: The expected frequency impact analysis module is used to: Performing LSTM-based time series coding on the future short-term power fluctuation prediction result to obtain a future short-term power fluctuation time series coding vector; Performing MLP-based data encoding on the real-time power grid data to obtain a real-time power grid data encoding vector; The future short-term power fluctuation time series coding vector and the power grid real-time data coding vector are fused and then input into a decoder-based frequency estimator to obtain the expected frequency impact assessment result.

8. The wind power grid-connected regulation system based on multivariate data analysis according to claim 7, characterized in that: The active frequency support strategy generation module includes an encoder and a decoder, wherein the encoder is used to perform structured encoding on the future short-term power fluctuation prediction result, the expected frequency impact assessment result, the frequency risk level and the real-time data of the power grid to obtain the future short-term power fluctuation prediction result structured encoding vector, the expected frequency impact assessment result structured encoding vector, the frequency risk level structured encoding vector and the real-time data of the power grid structured encoding vector; and the decoder is used to perform vector regression decoding on the cascade vector of the future short-term power fluctuation prediction result structured encoding vector, the expected frequency impact assessment result structured encoding vector, the frequency risk level structured encoding vector and the real-time data of the power grid structured encoding vector to obtain the virtual inertia coefficient, the reserved standby percentage, the power change rate limit value and the target active power adjustment value. The active frequency support strategy generation module is an end-to-end regression model, and its model parameters are obtained through a training process.

Images