Direct-drive wind turbine grid adaptive hybrid control method and system

Abstract

The application discloses a direct-drive wind turbine grid adaptability hybrid control method and system, belongs to the technical field of wind power generation control, collects the operation state parameters and grid disturbance signals of a direct-drive wind turbine in a wind power generation system, and constructs a disturbance perception state descriptor; the disturbance trajectory is mapped to a control strategy atlas space based on variational time series clustering, control domain characteristic coordinates are generated; a multi-strategy control atlas is called, a disturbance perception type control subgraph is constructed, path optimization and dynamic coupling are carried out, and a composite control vector is generated; the control vector is injected into a wind turbine converter control channel, the cooperative control of torque, current phase and reactive power support is realized; when a disturbance transition event is detected, the control structure reconstruction is automatically triggered, adaptive switching is realized; the method has the advantages of accurate disturbance identification, rapid control response and flexible strategy fusion, and improves the adaptability and stability of the wind turbine under complex grid conditions.

Description

Technical Field

[0001] This invention relates to the field of wind power generation control technology, specifically to a grid-adaptive hybrid control method and system for direct-drive wind turbines. Background Technology

[0002] Current wind power generation systems generally employ strategies such as maximum power point tracking (MPPT) and torque optimization, achieving high energy conversion efficiency under normal operating conditions. However, when the power grid experiences frequency changes, voltage fluctuations, or short-term disturbances, direct-drive wind turbines, lacking mechanical inertia and synchronization characteristics, struggle to provide stable inertial support and reactive power regulation. This can easily lead to increased system frequency deviation and widened voltage sags, thereby inducing the risk of grid disconnection.

[0003] Most existing inertia control and virtual synchronization control methods rely on fixed models or static response parameters, lacking the ability to dynamically identify disturbance modes and decouple intervention paths, making it difficult to adapt to the nonlinear evolution characteristics of power grid disturbances. Furthermore, existing control strategies typically employ a serial structure, with each control submodule lacking a fusion and coordination mechanism, leading to potential conflicts under multi-objective control tasks and reducing system response sensitivity and stability.

[0004] Therefore, it is urgent to break through the traditional technical framework based on single disturbance response or centralized control methods, and to construct a hybrid control method for direct-drive wind turbines that is guided by grid disturbance pattern clustering and integrates dynamic feature mapping and adaptive switching of control strategies. This method can realize the dynamic response of wind turbines to different types of grid disturbances through "target classification - control aggregation - path reconstruction", thereby improving their grid adaptability and grid support capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a grid-adaptive hybrid control method and system for direct-drive wind turbines to address the shortcomings in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a direct-drive wind turbine grid-adaptive hybrid control method, comprising: S1: Collect the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate. S2: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space through the variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; S3: Based on the characteristic coordinates of the control domain, call multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; S4: Perform path selection and dynamic coupling on each control strategy node in the control subgraph, construct a response path coordination matrix, so that different control strategies can achieve nonlinear superposition and temporal interleaving on the output side, and generate a composite control vector. S5: Apply the composite control vector to the wind turbine converter control channel, and perform modulation and injection according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability; S6: When a disturbance state transition occurs, the control mapping model automatically re-evaluates the control domain projection and triggers the response path reconstruction mechanism to complete the rapid adaptive switching of the wind turbine control structure.

[0007] Preferably, S1 includes: S11: Collect wind speed, fan speed, electromagnetic torque, output voltage and grid frequency change rate through a multi-dimensional asynchronous data interface, establish a high-frequency synchronous sampling buffer queue, and keep the timestamps of state parameters and disturbance signals aligned. S12: Input the collected data into the feature coupling module based on graph convolution structure, perform spatiotemporal deconstruction and channel compression on the dynamic correlation between parameters, and extract the disturbance sensitivity representation vector. S13: An embedded residual attention mechanism is adopted to fuse feature vectors and historical disturbance labels to achieve a joint representation of the current operating state and the power grid disturbance trend, generating a disturbance-aware state descriptor as the input of the disturbance-driven control mapping model.

[0008] Preferably, S2 includes: S21: The collected power grid disturbance signal sequence and the operating state descriptor are input into the disturbance-driven control mapping model. The model is based on a variational autoencoder structure to construct the disturbance evolution latent space and compress the non-stationary characteristics of the disturbance time series data. S22: Use a constrained multi-scale variational temporal clustering algorithm to classify the perturbation latent variable sequence and output the perturbation evolution trajectory cluster center and its probability distribution; S23: By projecting the features into the control strategy graph space through the center of the perturbation evolution cluster, the control domain feature coordinates corresponding to the perturbation are obtained, which serve as the input basis for subsequent control subgraph calls and path construction.

[0009] Preferably, S3 includes: S31: Establish a structured multi-strategy control graph, taking multiple atomic control strategies such as target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation as graph nodes, and constructing a graph edge matrix based on the cooperative relationship and conflict weight between strategies. S32: Based on the feature coordinates of the control domain, locate the center of the main control policy cluster corresponding to the feature coordinates through the graph retrieval function, and call its associated policy neighborhood nodes to generate a set of disturbance-related policy candidates. S33: Combining disturbance type labels with dynamic indicators of the current system state, a control response tensor is constructed using the joint response scoring function between policy nodes. A graph pruning operation is then performed to remove non-responsive edges and loosely coupled nodes, ultimately forming a disturbance-aware control subgraph.

[0010] Preferably, S4 includes: S41: Quantize and encode each atomic control strategy node in the control subgraph according to its strategy delay characteristics, control direction and output gain to form a multi-dimensional control attribute vector; S42: Construct a path evaluation network based on a dual attention guidance mechanism, adaptively weight the output conflict degree and response synergy between control nodes, and output a bidirectional influence graph between control strategies; S43: Based on the influence graph, perform path selection graph search, use multi-objective dynamic programming algorithm to extract the path combination with optimal coordination efficiency, and generate response path coordination matrix; S44: Based on the response path coordination matrix, nonlinear fusion is performed on the output of each control strategy node, and a composite control vector is generated using a time-interleaved gating structure, which serves as the final control instruction set injected into the wind turbine control execution module.

[0011] Preferably, S5 includes: S51: Input the composite control vector into the distributed modulation deconstruction module, and map it to the modulation parameter space of the converter according to the response priority and sensitivity coefficient of each strategy component in the vector, to construct three types of target control sets: torque modulation set, phase adjustment set and reactive power support set; S52: In the control set, a dependent scheduler based on the disturbance dynamic response window is introduced to analyze the cross-action relationship and phase delay feedback characteristics between each subset and to construct a cooperative modulation control spectrum. S53: By controlling the spectrum to drive the multi-path coupling injection mechanism, different control components are injected into the power channel, synchronous modulation unit and reactive power support path of the wind turbine converter respectively, and a dynamic complementary constraint model is introduced at the execution layer to achieve parallel stability of the three types of control objectives.

[0012] Preferably, S6 includes: S61: Construct a disturbance state variation detection module, which identifies disturbance transition events based on the continuous-time evolution trajectory of the disturbance-aware state descriptor and an adaptive sliding window entropy model based on the state jump rate and the disturbance phase inflection point. S62: After a transition event is detected, the control mapping model triggers a fast control domain relocation process, inputs the latest state descriptor into the lightweight control graph projection network, and re-estimates the control domain feature coordinates corresponding to the disturbance. S63: Compare the current control domain coordinates with the matching degree threshold of the area covered by the original control subgraph. If the difference exceeds the preset threshold, trigger the path evolution module to reselect policy nodes from the control graph and construct an updated response path coordination matrix. S64: Integrate the newly generated response path coordination matrix into the current control scheduling module to achieve rapid adaptive switching of the wind turbine control structure.

[0013] The present invention also provides a direct-drive wind turbine grid-adaptive hybrid control system, comprising: Data sensing and preprocessing module: collects the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate. Disturbance modeling and feature extraction module: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space through variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; Strategy Graph Management and Scheduling Module: Based on the characteristic coordinates of the control domain, it calls multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; Path planning and collaborative optimization module: performs path selection and dynamic coupling on each control strategy node in the control subgraph, constructs a response path collaborative matrix, enables different control strategies to achieve nonlinear superposition and temporal interleaving on the output side, and generates a composite control vector; Converter interface module: The composite control vector is applied to the wind turbine converter control channel, and modulation and injection are performed according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability; Adaptive reconfiguration module: When a disturbance state transition occurs, the control mapping model automatically re-evaluates the control domain projection and triggers the response path reconfiguration mechanism to complete the rapid adaptive switching of the wind turbine control structure.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a control mapping mechanism based on disturbance-aware state descriptors, combining variational temporal clustering and control strategy graph spatial projection to achieve accurate mapping from disturbance states to control strategy structures, effectively overcoming the limitations of traditional wind power control strategies such as "fixed logic, static switching, and response hysteresis." By introducing a control path coordination matrix and a temporal interleaving fusion mechanism, it can dynamically coordinate the output paths of different strategies, balancing response speed, output stability, and control coupling in multi-objective control, significantly improving the grid adaptability and support capabilities of wind turbine units in weak grids and complex disturbance scenarios.

[0015] 2. This invention innovatively introduces a disturbance transition detection and control structure-level reconfiguration mechanism, enabling rapid adaptive switching of the control system when a non-stationary transition occurs in the disturbance type. It eliminates the need for system re-initialization and possesses real-time evolution and high-response closed-loop capabilities. Compared to existing technologies, this invention demonstrates significant improvements in key performance indicators such as frequency support response time, voltage stabilization speed, and output stability. It exhibits strong engineering adaptability and application value, and is particularly suitable for complex power grid environments with a high proportion of renewable energy integration. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a mind map of the method of the present invention.

[0018] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, please refer to Figure 1 As shown in this embodiment, the direct-drive wind turbine grid-adaptive hybrid control method includes: S1: Collect the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate. S2: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space through the variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; S3: Based on the characteristic coordinates of the control domain, call multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; S4: Perform path selection and dynamic coupling on each control strategy node in the control subgraph, construct a response path coordination matrix, so that different control strategies can achieve nonlinear superposition and temporal interleaving on the output side, and generate a composite control vector. S5: Apply the composite control vector to the wind turbine converter control channel, and perform modulation and injection according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability. S6: When a disturbance state transition occurs, the control mapping model automatically re-evaluates the control domain projection and triggers the response path reconstruction mechanism to complete the rapid adaptive switching of the wind turbine control structure.

[0021] In this invention, the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point are first collected. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate.

[0022] Considering the heterogeneity of physical distribution and the asynchronous nature of data interfaces among various sensors and sampling channels, this invention adopts a multi-dimensional asynchronous data interface architecture in the acquisition structure. That is, parameters such as wind speed signal, mechanical rotation speed, electromagnetic torque, voltage waveform and power grid frequency variation are acquired through multiple independent channels. Each channel can operate independently based on different sampling frequencies and communication protocols.

[0023] To address the inconsistency in timing of asynchronous data, a high-frequency synchronous sampling buffer queue is constructed within the acquisition module. Each channel's acquired data is marked with a high-precision timestamp. The system aligns each sampling stream using a master clock signal and caches parameter samples from adjacent time periods in a central data structure, achieving time alignment and structured storage of state parameters and disturbance signals.

[0024] This cache queue features a multi-frame caching and sliding window structure, which ensures data continuity and stability during the initial and ongoing evolution stages of disturbances, providing high-quality input data for subsequent feature modeling.

[0025] After data acquisition and alignment are completed, the feature modeling stage begins. To comprehensively characterize the dynamic correlations between operating parameters and avoid the shortcomings of traditional linear feature fusion methods that cannot capture complex coupling relationships, this invention employs a Graph Convolutional Network (GCN) structure to construct the feature coupling module.

[0026] This module first constructs a dynamic feature map from the collected multi-source parameters, where each node represents a parameter channel such as wind speed, rotational speed, torque, voltage, and frequency change rate, and edges represent the coordinated changes between them per unit time. The edge weights are determined jointly by the time-series statistical covariance and the local disturbance gain coefficient, reflecting the coupling strength between the parameters.

[0027] In each round of convolution, the GCN module updates the representation of each parameter channel based on the adjacency matrix and node state information, enabling the transmission of higher-order structural information. After two layers of graph convolution, the system can extract potential perturbation cooperative features across parameters and time domains.

[0028] To further compress redundant information and focus on perturbation-sensitive features, this invention introduces a channel compression mechanism after graph convolution. Specifically, a learnable channel scoring function is used to assign weights to each feature channel, and sparse filtering is performed based on the scores to ultimately form a perturbation sensitivity representation vector. This vector captures the comprehensive feature representation of the system that best reflects the perturbation trend and dynamic interaction relationship.

[0029] While the aforementioned perturbation sensitivity representation vectors already possess strong perturbation feature recognition capabilities, they lack the ability to model the context of perturbation evolution trends and historical experience. Therefore, this invention designs an embedded residual attention mechanism fusion structure that deeply fuses the current representation vector with historical perturbation label information to generate a robust perturbation-aware state descriptor.

[0030] Specifically, the system's built-in perturbation label library is first invoked. This library contains the types of perturbations that occurred in historical operations and their response feature codes. The current representation vector is then input into the fusion module, semantically aligned with historical perturbation labels, and a multi-head attention mechanism is used to calculate its similarity weights across different perturbation semantics.

[0031] The attention result is used to reconstruct the current representation vector, aligning it with features enhanced in a direction similar to typical historical perturbation scenarios. Simultaneously, to prevent overfitting or information loss, the fusion module introduces a residual structure, concatenating the original representation with the attention-enhanced result along the channel dimension and compressing it into a fixed-length vector via a 1×1 convolution, forming the final perturbation-aware state descriptor.

[0032] This descriptor also contains: Characteristics of coupled disturbances in the current operating state; Similarity expression with historical perturbation patterns; It can be used to control the perturbation semantic coordinate features of the spatial projection of the map.

[0033] This state descriptor will then be passed as the core input to the disturbance-driven control mapping model, providing a precise basis for subsequent disturbance classification, control path generation, and control structure evolution.

[0034] This embodiment aims to perform dimensionality reduction encoding and evolution trajectory modeling on high-dimensional disturbance time-series data collected from wind power systems. To this end, the grid disturbance signal sequence output by the acquisition module and the operating state descriptor are jointly input into the disturbance-driven control mapping model.

[0035] The disturbance-driven control mapping model is built on a variational autoencoder structure and includes three main modules: encoder, latent space sampler, and decoder. Encoder structure: The perturbation signal sequence (such as voltage change rate, frequency change rate) is concatenated with the state descriptor and input into the encoding network. The network contains a dual-channel bidirectional gated recurrent unit (Bi-GRU) and a convolutional pooling module to extract the temporal evolution features and local change patterns of the perturbation. Latent space sampler: The encoder output is mapped to the latent mean vector μ and standard deviation vector σ of the perturbation. It is sampled through the reparameterization technique to generate the latent variable z of the perturbation evolution, which represents the state point of the perturbation in the latent space. Decoder structure: As a reconstruction constraint mechanism, z is decoded in reverse to fit the original perturbation sequence, and the model is optimized so that the latent variable z can truly express the structural characteristics of the perturbation.

[0036] This variational coding model has the ability to model non-stationary perturbations and implicit state transitions, avoiding the problem that traditional feature projection methods are insufficient in recognizing abrupt perturbations, and laying a structural foundation for subsequent perturbation classification and map mapping.

[0037] After obtaining the sequence of latent variables z of perturbation evolution, in order to extract its potential structure and evolution trend, this invention designs a constrained multi-scale variational temporal clustering algorithm for identifying and classifying perturbation patterns.

[0038] This clustering algorithm is based on the following key mechanism: Multi-scale feature construction: Multi-scale smoothing and dilatation sampling are performed on the latent variable z sequence to form a multi-layer feature view, which corresponds to short-period perturbations (such as instantaneous voltage drops) and long-period perturbations (such as frequency drift). Structural soft clustering mechanism: A Bayesian-Gaussian mixture model with prior constraints is used as the clustering basis. Physically heuristic perturbation features (such as voltage drop rate and inertial response window) are introduced as prior labels. The clustering probability distribution of the perturbation trajectory is obtained through expectation-maximization (EM) optimization. Perturbation trajectory cluster center extraction: After clustering, the center point vector of each perturbation cluster is calculated as a representative perturbation morphology code. The center point also carries the proportion probability of each type of perturbation cluster and the interpretable perturbation index.

[0039] This step enables the structural partitioning of perturbations in the latent space and outputs the centers of perturbation evolution trajectory clusters and their corresponding perturbation probability distributions, thereby forming a mapping from perturbation state to perturbation type.

[0040] Compared to traditional time-series classification methods such as K-means and DTW, this clustering mechanism introduces structural soft constraints and physical priors, which makes the disturbance identification results not only temporally consistent, but also reflect the severity of disturbances corresponding to actual control requirements.

[0041] To map the centers of the identified disturbance trajectory clusters to the control structure, this invention constructs a control strategy graph space in the control strategy management system. This graph space consists of multiple atomic control strategy nodes (such as target power tracking, virtual inertia support, short-time voltage stabilization, and reactive power allocation), and forms a topology graph structure through strategy coupling edges.

[0042] The center vector of the disturbance trajectory cluster will be used as the query input, and its mapping in the control map space will be completed through the following steps: Policy graph embedding modeling: Graph embedding training is performed on the control graph in advance to map policy nodes and their structural relationships to a unified vector space. Collaborative and exclusionary semantics between nodes are preserved through structure-preserving embedding algorithms (such as GraphSAGE or LINE). Feature alignment projection: The center of the disturbance trajectory cluster is used as the query vector. It is aligned with each policy node in the embedding space through similarity indicators such as cosine similarity or Mahalanobis distance to calculate the semantic fit between the disturbance and each control policy. Control domain feature coordinate extraction: Select the set of policy nodes with the highest fit and construct the vector coordinates of the perturbation control domain (i.e., control domain feature coordinates). These coordinates not only indicate the category of the optimal policy, but also contain the weight distribution and response priority information of the policy structure.

[0043] The characteristic coordinates of this control domain will serve as the input for subsequent control subgraph generation and path coordination matrix construction, enabling intelligent and structured matching of disturbance-driven control responses.

[0044] To structure and organize the controllable strategies of wind power systems, this invention first designs a multi-strategy control map to uniformly describe various control objectives and their synergistic relationships.

[0045] Each node in the graph represents an atomic control strategy, including but not limited to: Target power tracking; Virtual inertia support; Short-term voltage stabilization; Reactive power allocation strategy; Scalable strategy nodes include frequency response support, low voltage ride-through, and phase angle adjustment.

[0046] Each policy node is stored in a standard policy interface format, including attributes such as control objective, response time constant, applicable disturbance type, and master variable category.

[0047] The edges between nodes form the graph edge matrix, and the existence and weight of the edges are determined by the following two quantities: Strategy synergy: Quantifies whether two strategies can be executed in a positive synergistic manner in terms of timing and control. Control conflict weights: assess whether there are control resource conflicts or mutual exclusion of responses between the two strategies (such as competition for the same channel, current direction offsetting, etc.).

[0048] This graph structure can comprehensively express the coupling and constraint relationships between control strategies, providing a logical basis for the subsequent pruning and combination of control subgraphs.

[0049] In order to select the appropriate strategy set from the spectrum based on the actual disturbance state, this invention introduces a spectrum retrieval mechanism driven by control domain feature coordinates.

[0050] The steps are as follows: Feature coordinate input: The feature coordinates of the control domain are derived from the perturbation map projection output of step S2, including the semantics of the perturbation center, policy adaptation labels and local policy density distribution; Graph retrieval function call: Design the algorithm function F(Graph, Coord) →NodeSet, input the coordinates of the control domain, match the node cluster in the graph that has the largest fit with the coordinates, and define it as the center of the main control strategy cluster; Neighborhood expansion mechanism: With the master control policy cluster as the core, the neighborhood nodes are extended through the high cooperative edges in the graph to form a perturbation-related policy candidate set. This policy set is limited to having a cooperative foundation at the structural level.

[0051] This candidate set will serve as a preliminary strategy pool for control response planning, providing an input structure for response path generation and control tensor construction.

[0052] After obtaining the candidate policy set, it is necessary to further filter out the subset of policies that are most responsive to the current perturbation and have the highest execution value. To this end, this invention designs a control response tensor construction mechanism and a subgraph pruning algorithm, the steps of which are as follows: Disturbance label and status indicator fusion: The system simultaneously acquires disturbance type labels (such as frequency drop, voltage sag) and the current operating status of the wind turbine (such as torque margin, excitation current change rate, grid-side impedance information). Response scoring function definition: Calculate the joint response scoring function R(i, j) between candidate policy node pairs. This function comprehensively considers: The perturbation matching degree between nodes i and j; The node output shows the coverage of key control variables; Does the node combination have mutual exclusion for execution?

[0053] Control response tensor construction: The nodes R(i, j) are organized into a three-dimensional tensor structure to represent the effectiveness of the strategy combination in response under different control objective dimensions.

[0054] Graph clipping mechanism execution: Graph clipping operations are performed based on the control response tensor. Remove low-response edges (R(i,j) is below the threshold); Remove isolated nodes and strategies with low collaboration; The remaining structures are categorized hierarchically by strategy type.

[0055] The final retained subgraph is the disturbance-aware control subgraph, which contains a set of control strategy nodes with the highest response efficiency under the current disturbance conditions and their cooperative path relationships, serving as the structural input for subsequent path optimization and control fusion.

[0056] The control subgraph contains multiple atomic control strategy nodes, each representing a specific functional module (such as MPPT, virtual inertia support, etc.). To achieve comparability and structured input, this invention quantizes and encodes these nodes according to three key control attributes, constructing a multi-dimensional control attribute vector, specifically including: Policy latency characteristic: This represents the average response time required for the policy to take effect from triggering to control quantity taking effect, in milliseconds, and is obtained through experimental system identification.

[0057] Control direction: refers to the main active variable dimension of the control strategy, such as whether it acts on active / reactive power, direct-axis / quadrature-axis or converter grid-side / machine-side channels, and is represented by a one-hot vector.

[0058] Output gain coefficient: represents the control strategy's ability to respond to the magnitude of the main control variable (such as the amplification factor of Δ wind speed to Δ torque), and is a normalized floating-point scalar.

[0059] The three attributes are concatenated to form a unified control attribute vector, which is used to describe the response behavior of each strategy node and the system action path, providing quantitative input for subsequent modeling of strategy collaboration relationships.

[0060] In order to establish accurate output coordination relationships between policy nodes and evaluate whether they can form an efficient collaborative path, this invention designs a path evaluation network based on a dual attention mechanism to model the conflict and collaboration between policies respectively.

[0061] The conflict attention module utilizes the angular relationship between the node control direction and the output gain to determine whether policies exhibit output mutual exclusion or resource contention. For example, when two policies act on the same converter channel but in opposite directions, their conflict weight increases. This module outputs a conflict attention matrix, Aconflict.

[0062] The Collaborative Attention module analyzes the overlap of policy response times, the complementarity of control directions, and the interoperability of state variables to evaluate their collaborative potential. This module outputs the collaborative attention matrix Acooperate.

[0063] By combining the two attention matrices mentioned above, a bidirectional influence graph of the control policy is constructed, where edge weights represent the potential for effective combinations between policies. This graph structure will serve as input to the path search algorithm.

[0064] On the constructed two-way influence graph of the control strategy, this invention uses a multi-objective dynamic programming algorithm for path search and optimization, with the goal of selecting a set of path combinations that have the highest coordination efficiency and the lowest conflict risk.

[0065] Objective function setting: Taking into account the following three objectives: Maximize the collaborative weights between nodes (from Acooperate); Minimize the impact of output conflicts (derived from Aconflict); The ability of the optimized path to cover the target variables of the system (such as the coverage of the three types of targets: power, phase, and reactive power).

[0066] Graph search and path combination extraction: A graph search algorithm based on Pareto optimality is used to output a set of candidate combinations of control paths, and to perform path deduplication and redundant node removal.

[0067] Response path coordination matrix generation: The final selected path is mapped into a response path coordination matrix. Each row of the matrix represents a control node, and each column represents the actual strength and rhythm window of its effect on the control target.

[0068] To achieve multi-strategy fusion output after path coordination, this invention introduces a time-interleaved gating structure to dynamically fuse the control strategy output.

[0069] Fusion modulation structure: The output signal of each policy node first passes through its corresponding weight modulator in RPCM, and then passes through the gating module to determine its activation timing (leader period, steady state period or delay period).

[0070] Interleaved control window settings: The activation windows of different strategies are interleaved to ensure that the control output rhythm is staggered and to prevent signal overlap from causing system oscillation or response delay.

[0071] Composite control vector generation: All strategy outputs are time-series superimposed and amplitude-fused to form a composite control vector, which is then injected into the wind turbine control execution module to adjust the output torque, current phase, and reactive power support.

[0072] The composite control vector is an integrated control command generated in step S4 based on a time-interleaving mechanism, containing multiple policy components (such as MPPT output, virtual inertia output, reactive power regulation command, etc.). Since these policies differ in their direction of action, time response, and output variable types, they need to be structured, deconstructed, and mapped.

[0073] To this end, the present invention constructs a distributed modulation and deconstruction module, which processes the composite control vector input according to the following logic: Policy component parsing: Each component in the control vector is classified according to its policy identifier (such as node number, control variable label) to identify its target and dynamic attributes.

[0074] Response priority extraction: Read the response priority and disturbance sensitivity coefficient of each strategy component in the path coordination matrix, and construct a priority weight table.

[0075] Modulation parameter space mapping: The above components are mapped to the converter control channels corresponding to the three types of control objectives, thus constructing the three types of target control sets: Torque modulation set: includes strategies for directly controlling the electromagnetic torque of the generator; Phase adjustment set: includes strategies that affect the synchronization of current phase angle and voltage; Reactive power support set: includes strategies for regulating reactive power output and grid-side voltage support.

[0076] The above three sets constitute the main control objective group of the control system in the disturbance response phase, providing a deconstruction basis for subsequent multi-objective scheduling.

[0077] To achieve coordinated scheduling among multiple modulators and avoid mutual exclusion conflicts between control channels, this invention proposes a dependent scheduler that constructs a timing and phase dependency structure among control objectives, ultimately generating a cooperative modulation control spectrum.

[0078] Its core steps include: Dynamic response window extraction: Set a dynamic modulation window based on the current disturbance type (such as frequency drop, voltage fluctuation) to identify the key response stages (leading response stage, reconstructed steady state stage, recovery period stage, etc.) of each control target in the disturbance evolution process.

[0079] Dependency computation logic: Lateral dependence: Determine whether there is amplitude coupling between torque modulation and phase adjustment in the early stage (such as the synchronous relationship between electromagnetic torque and current angle); Vertical feedback dependence: Detect the feedback effect of reactive power support strategy on torque response and establish feedback edges.

[0080] Modulation graph generation: Construct a dependency graph structure among three types of control sets, where nodes represent control subsets and edges represent temporal dependencies or modulation conflict relationships. The graph structure is stored in a graph database and includes time stamps and execution priority information.

[0081] The collaborative graph, as the core organizational structure for controlling the execution path, is used to guide the rhythm control and intervention sequencing of multi-target injection mechanisms.

[0082] After the control graph is constructed, the system schedules each control subset according to the graph structure and executes the multi-path coupling injection mechanism through the converter interface to realize the execution of control commands.

[0083] Channel partitioning and injection interface: Injecting torque modulation sets into the power path of the generator-side converter; The phase adjustment set is injected into the grid-side synchronization control unit to adjust the phase angle and phase-locked loop parameters; The reactive power support is injected into the current component control path to adjust the q-axis component and support the voltage.

[0084] Introduction of dynamic complementary constraint model: To ensure that the three types of control objectives do not interfere with each other, the system constructs a dynamic complementary constraint model at the converter control layer, the functions of which include: Limit the power regulation rate to prevent surges; Dynamically adjust PI parameters to adapt to timing overlap in the control channels; Short-term complementary windows are set based on predicted disturbance trends to achieve "off-peak execution" of control channels.

[0085] Parallel control stability guarantee: Based on the scheduling rhythm of the modulation spectrum, the system implements non-blocking parallel control at the hardware level, enabling the three types of targets to complete synchronous modulation in milliseconds, thereby improving the response efficiency and operational stability of the wind turbine converter in multi-target scenarios.

[0086] To achieve dynamic reconfiguration of the control structure, it is essential to first accurately identify whether structural abrupt changes occur in the perturbation state. This invention constructs a perturbation state variation detection module to continuously monitor the evolution trajectory of the perturbation state and identify perturbation transition events.

[0087] The module takes as input a sequence of disturbance-aware state descriptors, derived from the high-dimensional state-encoded data generated in steps S1 and S2. The detection method is as follows: Calculate the rate of change of the Euclidean distance between the current state descriptor and the previous time descriptor. If the rate continuously exceeds the set threshold within a short time window, it is considered that there is a significant perturbation structure jump.

[0088] Principal component analysis (PCA) is performed on the evolution trajectory of the state descriptor to reduce dimensionality and obtain the main perturbation change directions. The sliding window function entropy method is used to analyze the abrupt change points in the first derivative direction of the trajectory and identify the "phase inflection points" of the perturbation trend.

[0089] By combining the jump rate and phase transition characteristics, an adaptive sliding window entropy judgment model is constructed. By comprehensively considering the intensity of disturbance fluctuations, the degree of directional change, and the persistence, it is determined whether the disturbance has evolved from the same type of disturbance to a new type of disturbance, and then it is determined whether it is a transition event.

[0090] When the above model determines that the current disturbance state has transitioned, the system triggers the control mapping relocation process.

[0091] Once a disturbance transition event is detected, the original control path may no longer be suitable for the current disturbance characteristics, thus requiring a re-alignment of the control graph space strategy. To address this, this invention designs a lightweight control graph projection network to achieve rapid relocation of the disturbance state.

[0092] Extract the state descriptor of the most recent moment from the current disturbance sensing module, which contains the main dynamic response characteristics and state variable distribution of the new disturbance.

[0093] The state descriptor is input into a lightweight graph embedding network (such as Tiny-GCN or Distilled GraphSAGE) to quickly compute the corresponding control domain feature coordinates (i.e., the semantic center of the control policy) in the control policy graph space.

[0094] Compared to the full graph projection network used in step S2, the lightweight model reduces the number of network layers and embedding dimensions, adapting to the stringent response speed requirements of real-time scenarios, with a typical response latency of less than 20ms. The new control domain feature coordinates output in this step are used to determine whether structural reconstruction of the policy path is needed.

[0095] After the control domain coordinates are updated, it is necessary to determine whether the original control subgraph still fits the current disturbance state. To this end, this invention introduces a matching degree threshold determination mechanism and a path evolution module to dynamically update the control structure.

[0096] Calculate the cosine similarity and policy coupling coverage between the current control domain coordinates and the policy region covered by the original control subgraph to determine their structural matching degree.

[0097] If the matching degree is lower than the reconstruction threshold set by the system (e.g., lower than 0.75), the original strategy path is considered unsuitable, and the system will trigger path evolution.

[0098] Perform a policy retrieval and path construction process similar to S3–S4, reselect a policy set that is adapted to the current disturbance state in the control policy graph, and generate a new response path coordination matrix for subsequent control fusion.

[0099] After the newly generated response path coordination matrix is ​​completed, the system needs to realize real-time replacement of the control structure and switching of the execution link. This invention designs a structure-level control switching mechanism to achieve seamless switching of the control system structure.

[0100] The control and scheduling module adopts a modular structure, and each strategy node and control chain can be loaded as a "control plug-in" on demand. The control strategy corresponding to the new path coordination matrix will be dynamically and hot-swapped without interrupting system operation, ensuring continuity.

[0101] The system records control state variables (such as initial integral values, delay buffers, etc.) and migrates them to the new control node to avoid state loss or response oscillation during the switching process.

[0102] To address output jitter caused by instantaneous control chain switching, the system introduces short-time smoothing filtering and gradual adjustment strategies to maintain continuous stability of output power, voltage phase, and reactive power support.

[0103] Ultimately, by rapidly deploying the new path structure and synchronously updating the execution chain, the wind turbine control system achieves structural-level adaptive switching under disturbance transition conditions.

[0104] Example 2: To verify the disturbance-driven control mapping model, control graph coordination mechanism, and structure-level dynamic switching capability described in this invention, this example was experimentally verified on a hardware-in-the-loop simulation platform and compared with the traditional dual-closed-loop vector control method.

[0105] Simulation platform: Based on MATLAB / Simulink + OPAL-RT real-time controller, a hardware-in-the-loop simulation architecture for wind power systems is constructed; Wind turbine model: A typical 2MW permanent magnet direct-drive wind turbine model is adopted, with complete electromechanical coupling dynamics and converter model; Power grid model: weak power grid scenario (short-circuit ratio SCR ≈ 1.5), supporting voltage step disturbance and frequency sag simulation; The present invention proposes the following solutions: introducing disturbance sensing control mapping + multi-strategy graph + path coordination matrix + structure-level switching mechanism; Comparison scheme: Adopting a parallel control structure of traditional MPPT + voltage / frequency response function module, with disturbance-free identification and strategy adaptation.

[0106] The experiment sets up two typical disturbance scenarios to examine the response speed, stability, and grid support capability of the control system: Scene 1 sudden drop in power grid frequency The frequency dropped from 50Hz to 48.5Hz, lasting for 500ms. Scene 2 instantaneous voltage drop The grid connection point voltage dropped from 1.0 pu to 0.6 pu and recovered within 200ms. The following four indicators are used to measure control performance: Frequency support response time (T1): The time taken from the occurrence of frequency disturbance to the rapid increase of the fan output power; Voltage recovery time (T2): The time required for the reactive power injected by the fan to recover to 0.95 pu after a voltage drop; System oscillation amplitude (A1): The maximum peak deviation of the fan output current and torque after the disturbance; Structure switching delay (T3): The average system delay for detecting disturbance transitions to complete the switching on the control path.

[0107] Comparison of experimental results: <![CDATA[Frequency support response time T1]]> 92 ms 173 ms ↑46.8% <![CDATA[Voltage recovery time T2]]> 124 ms 233 ms ↑46.8% <![CDATA[Output oscillation amplitude A1]]> ±4.2% ±9.7% ↓56.7% <![CDATA[Structure switching delay T3]]> 38 ms Automatic switching is not supported. — Graphical data shows that after a frequency drop, the control system of this invention can quickly activate the "virtual inertia + phase guidance" path and significantly improve the active power output of the wind turbine within less than 100ms. In the case of voltage drop, the control system can quickly switch to a control path based on the "reactive power support + current limiting" strategy through disturbance identification, effectively suppressing voltage fluctuations and restoring system stability.

[0108] Experimental results demonstrate that the hybrid control strategy described in this invention has the following advantages in dealing with power grid disturbances: it can complete high-precision identification and strategy path adjustment in the early stages of disturbances; it enables on-demand combination of control strategies rather than fixed calls, avoiding output conflicts; the control structure has adaptive reconfiguration capabilities, ensuring control continuity; and its response speed and system stability are significantly better than traditional schemes, making it suitable for weak power grids and environments with multiple disturbances.

[0109] Example 3, please refer to Figure 2As shown, the direct-drive wind turbine grid-adaptive hybrid control system described in this embodiment includes: Data sensing and preprocessing module: collects the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate. Disturbance modeling and feature extraction module: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space through variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; Strategy Graph Management and Scheduling Module: Based on the characteristic coordinates of the control domain, it calls multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; Path planning and collaborative optimization module: performs path selection and dynamic coupling on each control strategy node in the control subgraph, constructs a response path collaborative matrix, enables different control strategies to achieve nonlinear superposition and temporal interleaving on the output side, and generates a composite control vector; Converter interface module: The composite control vector is applied to the wind turbine converter control channel, and modulation and injection are performed according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability; Adaptive reconfiguration module: When a disturbance state transition occurs, the control mapping model automatically re-evaluates the control domain projection and triggers the response path reconfiguration mechanism to complete the rapid adaptive switching of the wind turbine control structure.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A grid-adaptive hybrid control method for direct-drive wind turbines, characterized in that: include: S1: Collect the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, turbine speed, electromagnetic torque, output voltage, and grid frequency change rate. S1 includes: S11: Collect wind speed, fan speed, electromagnetic torque, output voltage and grid frequency change rate through a multi-dimensional asynchronous data interface, establish a high-frequency synchronous sampling buffer queue, and keep the timestamps of state parameters and disturbance signals aligned. S12: Input the collected data into the feature coupling module based on graph convolution structure, perform spatiotemporal deconstruction and channel compression on the dynamic correlation between parameters, and extract the disturbance sensitivity representation vector. S13: An embedded residual attention mechanism is used to fuse the disturbance sensitivity representation vector and historical disturbance labels to achieve a joint representation of the current operating state and the power grid disturbance trend, generating a disturbance-aware state descriptor as input to the disturbance-driven control mapping model. The disturbance-aware state descriptor includes: coupled disturbance features under the current operating state; similarity expression with historical disturbance patterns; and disturbance semantic coordinate features that can be used for control map spatial projection. S2: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space using a variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; S2 includes: S21: The collected power grid disturbance signal sequence and the disturbance sensing state descriptor are input into the disturbance-driven control mapping model. The disturbance-driven control mapping model constructs the disturbance evolution latent space based on the variational autoencoder structure to compress the non-stationary characteristics of the disturbance time series data. S22: The perturbation latent variable sequence is classified into perturbation morphology using a constrained multi-scale variational temporal clustering algorithm, and the perturbation evolution trajectory cluster center and its probability distribution are output. The perturbation latent variable sequence represents the state point of the perturbation in the perturbation evolution latent space. S23: By projecting the features of the perturbation evolution trajectory cluster center into the control strategy map space, the control domain feature coordinates corresponding to the perturbation are obtained, which serve as the input basis for subsequent control subgraph calls and path construction; S3: Based on the characteristic coordinates of the control domain, call multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; S4: Perform path selection and dynamic coupling on each control strategy node in the disturbance-aware control subgraph, construct a response path coordination matrix, so that different control strategy nodes can achieve nonlinear superposition and temporal interleaving on the output side, and generate a composite control vector. S5: Apply the composite control vector to the wind turbine converter control channel, and perform modulation and injection according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability; S6: When a disturbance state transitions, the control disturbance drives the control mapping model to automatically re-evaluate the control domain projection and trigger the response path reconstruction mechanism to complete the rapid adaptive switching of the wind turbine control structure.

2. The direct-drive wind turbine grid-adaptive hybrid control method according to claim 1, characterized in that: S3 includes: S31: Establish a structured multi-strategy control graph, taking multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, as graph nodes, and constructing a graph edge matrix based on the cooperative relationship and conflict weights between strategies. S32: Based on the feature coordinates of the control domain, the center of the main control strategy cluster corresponding to the feature coordinates is located through the graph retrieval function. With the center of the main control strategy cluster as the core, the neighboring nodes are extended through the high cooperative edges in the graph to form a set of perturbation-related strategy candidates. S33: Combining the disturbance type label with the dynamic index of the current system state, a control response tensor is constructed using the joint response scoring function between policy nodes. A graph pruning operation is performed to remove low-responsibility edges and low-coupling nodes, ultimately forming a disturbance-aware control subgraph. The low-responsibility edges are those whose joint response scoring function is below a threshold.

3. The direct-drive wind turbine grid-adaptive hybrid control method according to claim 2, characterized in that: S4 includes: S41: Quantize and encode each atomic control strategy node in the disturbance-aware control subgraph according to its strategy delay characteristics, control direction and output gain to form a multi-dimensional control attribute vector; S42: Construct a path evaluation network based on a dual attention guidance mechanism, adaptively weight the output conflict degree and response synergy between control policy nodes, and output a bidirectional influence graph between control policy nodes; S43: Based on the bidirectional influence graph, perform path selection graph search, use multi-objective dynamic programming algorithm to extract the path combination with optimal coordination efficiency, and generate response path coordination matrix; S44: Based on the response path coordination matrix, nonlinear fusion is performed on the output of each control strategy node, and a composite control vector is generated using a time-interleaved gating structure, which serves as the final control instruction set injected into the wind turbine control execution module.

4. The direct-drive wind turbine grid-adaptive hybrid control method according to claim 1, characterized in that: S5 includes: S51: Input the composite control vector into the distributed modulation deconstruction module, and map it to the modulation parameter space of the converter according to the priority and dynamic sensitivity coefficient of each strategy component in the composite control vector, and construct three types of target control sets: torque modulation set, phase adjustment set and reactive power support set; S52: In the control set, a dependent scheduler based on the disturbance dynamic response window is introduced to analyze the cross-action relationship and phase delay feedback characteristics between each subset and to construct a cooperative modulation control spectrum. S53: By driving a multi-path coupling injection mechanism through a collaborative modulation control spectrum, different control components are injected into the power channel, synchronous modulation unit and reactive power support path of the wind turbine converter respectively. A dynamic complementary constraint model is introduced at the execution layer to achieve parallel stability of the three types of control objectives. The dynamic complementary constraint model includes limiting the power regulation rate, dynamically adjusting the PI parameter and setting a short-term complementary window based on the predicted disturbance trend to achieve peak-shifting execution of the control channel.

5. The direct-drive wind turbine grid-adaptive hybrid control method according to claim 3, characterized in that: S6 includes: S61: Construct a disturbance state variation detection module, which identifies disturbance transition events based on the continuous-time evolution trajectory of the disturbance-aware state descriptor and an adaptive sliding window entropy model based on the state jump rate and the disturbance phase inflection point. S62: After detecting a transition event, the control disturbance drives the control mapping model to trigger a fast control domain relocation process. The latest disturbance-aware state descriptor is input into the lightweight control policy graph projection network to re-estimate the control domain feature coordinates corresponding to the disturbance. S63: Compare the current control domain coordinates with the matching degree threshold of the area covered by the original control subgraph. If the difference exceeds the preset threshold, trigger the path evolution module to reselect policy nodes from the control policy graph and construct an updated response path coordination matrix. S64: Integrate the newly generated response path coordination matrix into the current control scheduling module to achieve rapid adaptive switching of the wind turbine control structure.

6. A direct-drive wind turbine grid-adaptive hybrid control system, used to implement the direct-drive wind turbine grid-adaptive hybrid control method according to any one of claims 1-5, characterized in that: include: Data sensing and preprocessing module: collects the operating status parameters of the direct-drive wind turbine in the wind power generation system and the grid disturbance signal at the grid connection point. The operating status parameters include wind speed, wind turbine speed, electromagnetic torque, output voltage and grid frequency change rate. Disturbance modeling and feature extraction module: Based on the disturbance signal input disturbance-driven control mapping model, the disturbance evolution trajectory is projected onto the control strategy map space through variational time-series clustering algorithm to obtain the control domain feature coordinates corresponding to the disturbance; Strategy Graph Management and Scheduling Module: Based on the characteristic coordinates of the control domain, it calls multiple atomic control strategies, including target power tracking, virtual inertia support, short-time voltage stabilization and reactive power allocation, from the preset multi-strategy control graph to construct a disturbance-aware control subgraph; Path planning and collaborative optimization module: performs path selection and dynamic coupling on each control strategy node in the disturbance-aware control subgraph, constructs a response path collaborative matrix, and enables different control strategy nodes to achieve nonlinear superposition and temporal interleaving on the output side to generate a composite control vector; Converter interface module: The composite control vector is applied to the wind turbine converter control channel, and modulation and injection are performed according to priority and dynamic sensitivity coefficient to achieve coordinated optimization control of wind turbine output torque, current phase and reactive power support capability; Adaptive reconfiguration module: When a disturbance state transitions, the control disturbance drives the control mapping model to automatically re-evaluate the control domain projection and trigger the response path reconfiguration mechanism, thus completing the rapid adaptive switching of the wind turbine control structure.

Images