Artificial intelligence based multi-modal wind turbine electromechanical transient modeling method

Abstract

The application discloses a multi-modal wind turbine electromechanical transient modeling method based on artificial intelligence, relates to the technical field of new energy power generation modeling, and comprises the following steps: using a causal discovery algorithm to perform causal correlation mining and dependent path identification on a standardized space-time data cube, constructing a causal topology graph, and performing cross-modal feature fusion through a space-time graph attention network to generate a high-order feature tensor; dividing the high-order feature tensor into a meta-learning task pool according to different machine types and environmental conditions, and training a cross-machine type parameterization framework by using a double-layer optimization strategy; constructing a composite working condition generator based on the trained cross-machine type parameterization framework, and performing constraint through a causal regularization loss function to form an extended test working condition set; and using a multi-modal deep reinforcement learning algorithm to optimize control parameters of the extended test working condition set to obtain optimized control parameters. The application lays a foundation for realizing electromechanical transient modeling with high robustness and high generalization ability through accurate causal modeling and cross-modal fusion.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation modeling technology, and in particular to a multimodal wind turbine electromechanical transient modeling method based on artificial intelligence. Background Technology

[0002] With the continuous increase in the penetration rate of new energy power systems, electromechanical transient modeling technology for wind turbines has become a core support for ensuring the stable operation of the power grid. Traditional modeling methods are mainly based on the construction of physical mechanism models and parameter tuning. This involves building detailed electromagnetic transient models of doubly-fed or direct-drive wind turbines and verifying parameters using hardware-in-the-loop test data from the controller. In recent years, data-driven modeling methods have gradually emerged, employing techniques such as time-domain waveform matching and frequency-domain impedance fitting to perform black-box modeling of the dynamic response characteristics of the turbine. The IEC 61400-27 series of standards developed by the International Electrotechnical Commission (IEC) defines a general electromechanical model framework for typical wind turbines, providing a basic modeling paradigm for the industry. However, in actual engineering, the controller strategies of different manufacturers vary significantly, and existing methods still face severe challenges when dealing with multiple turbine models and operating conditions. Specifically, these challenges manifest as two major technical bottlenecks: low efficiency of cross-modal data fusion and insufficient control parameter migration capabilities.

[0003] In existing technologies, modeling methods based on multimodal data generally employ feature-level fusion strategies, using manually designed time-frequency domain feature extraction rules to perform correlation analysis on electrical, mechanical, and environmental data of generating units. For example, wavelet packet decomposition is used to extract harmonic features of voltage drop transient processes, and a random forest classification model is constructed by combining power curve data. However, these methods have limitations. Manual feature engineering is insufficient to effectively capture the nonlinear causal relationships between multi-source heterogeneous data, especially since the coupling mechanism between electrical transient processes and unit mechanical dynamics during grid faults is not fully modeled, leading to a sharp decline in generalization performance under novel topologies or unknown control strategies. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based multimodal wind turbine electromechanical transient modeling method to solve the technical problem of missing feature correlation in traditional methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides an artificial intelligence-based method for electromechanical transient modeling of multimodal wind turbines. The method includes: collecting multimodal data from the wind turbine and preprocessing it to generate a standardized spatiotemporal data cube; using a causal discovery algorithm to mine causal relationships and identify dependency paths within the standardized spatiotemporal data cube, constructing a causal topology graph, and performing cross-modal feature fusion through a spatiotemporal graph attention network to generate a high-order feature tensor; dividing the high-order feature tensor into meta-learning task pools according to different turbine models and environmental conditions, and training a cross-model general parameterization framework using a two-layer optimization strategy; constructing a composite operating condition generator based on the trained cross-model general parameterization framework, constraining it using a causal regularization loss function to form an extended test operating condition set; using a multimodal deep reinforcement learning algorithm to optimize the control parameters of the extended test operating condition set to obtain optimized control parameters; and using the optimized control parameters to construct a digital twin virtual operating condition running instance, performing dynamic time warping verification, and generating a standardized verification report.

[0008] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps for generating the standardized spatiotemporal data cube are as follows:

[0009] Collect voltage waveforms, current waveforms, controller register signals, and environmental monitoring data to obtain multimodal data;

[0010] The multimodal data is aligned according to timestamps, and the time-aligned multimodal data sequences are processed by missing value imputation, outlier removal and noise suppression to obtain a cleaned high-quality dataset.

[0011] The cleaned high-quality dataset is uniformly resampled according to modality type and sampling frequency, and then normalized and standardized by unit to generate a standardized numerical matrix.

[0012] The standardized numerical matrix is ​​organized into a three-dimensional structure according to time steps, spatial nodes and modal dimensions to generate a standardized spatiotemporal data cube.

[0013] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps of using a causal discovery algorithm to perform causal association mining and dependency path identification on a standardized spatiotemporal data cube to construct a causal topology graph are as follows.

[0014] The standardized spatiotemporal data cube is expanded into multivariate time series samples. By using conditional independence tests and time delay analysis, potential causal relationships between state variables in the multivariate time series samples are identified.

[0015] The identified potential causal relationships are constructed as directed edges, and all state variables are used as nodes to generate a preliminary causal graph structure.

[0016] A structural optimization algorithm is used on the initial causal graph structure to remove false associations and redundant paths, resulting in a sparse causal topology graph.

[0017] Based on the node dependencies and temporal order in the sparse causal topology graph, the directionality and strength of the edges are labeled to form a causal topology graph.

[0018] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps for generating the high-order feature tensor are as follows:

[0019] The causal topological graph is encoded as a graph structure tensor, a graph representation processing structure is constructed, and the association information of nodes in the causal topological graph is fused using an attention mechanism to extract multi-scale representations that reflect the structural hierarchical features and generate a unified modal representation.

[0020] By utilizing the temporal attention mechanism, the uniform modal representation is weighted and aggregated within a historical time window to obtain a temporal evolution feature representation;

[0021] The unified modal representation and the temporal evolution feature representation are fused through feature concatenation to generate a high-order feature tensor.

[0022] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the steps of dividing the high-order feature tensor into meta-learning task pools according to different turbine models and environmental conditions, and training a cross-turbine general parameterization framework using a two-layer optimization strategy are as follows:

[0023] The high-order feature tensors are classified according to the device type and environmental conditions, and divided into multiple meta-learning task pools. Based on the samples in the meta-learning task pools, the global parameters of the cross-device general parameterization framework are initialized.

[0024] In each meta-learning task pool, task-specific data is used to perform local task optimization on the cross-machine general parameterization framework to obtain task adaptation parameters.

[0025] The task adaptation parameters from all meta-learning task pools are aggregated, and a two-layer optimization strategy is used to update the initialized global parameters to obtain the optimized global parameters.

[0026] Repeat the iterative process of local task optimization and global parameter update until the cross-device general parameterization framework training converges, and output the trained cross-device general parameterization framework.

[0027] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps for forming the extended test condition set are as follows:

[0028] Based on the trained cross-model general parameterization framework, structural parameters and representation vectors are extracted from the high-order feature tensor and input into the composite working condition generator for initialization processing to obtain the initial structural parameters of the composite working condition generator.

[0029] The initial structural parameters of the composite working condition generator are preheated and trained using real working condition samples. After training, a composite working condition generator with preliminary working condition construction capability is obtained, which generates preliminary working condition samples. Based on the causal relationship between nodes in the causal topology graph, a causal regularization loss function is constructed.

[0030] Based on the initial working condition samples and the constructed causal regularization loss function, the generation loss and the causal regularization loss based on the causal consistency metric are calculated using an adversarial training method. The parameters of the composite working condition generator are jointly optimized to obtain the optimized composite working condition generator and output the working condition samples.

[0031] The output working condition samples are subjected to causal consistency tests and distribution characteristic screening to form an extended test condition set.

[0032] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps for obtaining the optimized control parameters are as follows:

[0033] By utilizing an extended test case set, a multimodal environment state representation is constructed, and the policy gradient method is used to evaluate the policy of the multimodal environment state representation to generate control parameter policies.

[0034] The control parameter strategy is input into the simulation environment corresponding to the extended test condition set, and the simulation response process is monitored. Control effect data is collected, and feedback rewards are calculated based on preset performance indicators.

[0035] Based on feedback rewards, an optimization algorithm is used to adjust the control parameter strategy, and the training is repeated until the preset performance index is reached to obtain the optimized control parameters.

[0036] As a preferred embodiment of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence described in this invention, the specific steps for generating the standardized verification report are as follows:

[0037] By expanding the test case set and optimizing control parameters, a digital twin virtual test case running instance is constructed.

[0038] Based on the optimized control parameter input digital twin virtual operating condition instance, the time series response data of the operating instance is collected and dynamically time-warped with multimodal data to obtain aligned time series data;

[0039] Based on aligned time series data, the root mean square error and dynamic time warping distance are used to calculate the error index and dynamic performance deviation between the digital twin virtual operating condition instance and the extended test condition set. The response consistency of the digital twin virtual operating condition instance is evaluated, and a standardized verification report is generated.

[0040] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence as described in the first aspect of the present invention.

[0041] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence as described in the first aspect of the present invention.

[0042] The beneficial effects of this invention are as follows: By employing a causal discovery algorithm to mine causal associations and identify dependency paths in a standardized spatiotemporal data cube, a deep analysis of the potential causal relationships within multimodal spatiotemporal data is achieved. This helps reveal the true causal dependency structure between state variables, avoiding the problems of causal confusion and redundant information interference in traditional correlation analysis. By constructing a causal topology graph and using a spatiotemporal graph attention network to weightedly fuse the association information of nodes in the causal structure, efficient integration of cross-modal features and generation of high-order feature tensors are achieved. Through accurate causal modeling and cross-modal fusion, a solid foundation is laid for achieving highly robust and generalizable electromechanical transient modeling, effectively improving accuracy and adaptability. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of the multimodal wind turbine electromechanical transient modeling method based on artificial intelligence in this invention.

[0045] Figure 2 This is a flowchart of the standardized spatiotemporal data cube generation process in this invention.

[0046] Figure 3 This is a flowchart of the causal topology graph construction in this invention.

[0047] Figure 4 This is a flowchart of the training of the cross-model universal parameterization framework in this invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This embodiment provides an artificial intelligence-based method for electromechanical transient modeling of multimodal wind turbines, including the following steps:

[0052] S1. Collect multimodal data from wind turbine generators and preprocess it to generate a standardized spatiotemporal data cube.

[0053] S1.1 Acquire voltage waveforms, current waveforms, controller register signals, and environmental monitoring data to obtain multimodal data.

[0054] Specifically, the acquisition of voltage waveform data includes: sampling voltage signals at key nodes of the wind turbine using voltage sensors, with a sampling frequency of several kilohertz per second, and recording the voltage waveform data in real time; the acquisition of current waveform data includes: sampling the current of each phase of the wind turbine using current sensors, with a sampling frequency of several kilohertz per second, and recording the current waveform data in real time; the acquisition of controller register signals includes: reading the status signals and control parameters in the controller register through the controller communication interface, with a reading frequency of several tens of times per second, and acquiring the controller register signals; the acquisition of environmental monitoring data includes: measuring environmental parameters such as wind speed, temperature, and humidity using environmental sensors, with a sampling frequency of once per minute, and recording the environmental monitoring data; finally, the voltage waveform data, current waveform data, controller register signals, and environmental monitoring data are synchronized according to timestamps to obtain multimodal data.

[0055] S1.2 Align the multimodal data according to the timestamp, and perform missing value imputation, outlier removal and noise suppression on the time-aligned multimodal data sequence to obtain a cleaned high-quality dataset.

[0056] Specifically, the voltage waveform data, current waveform data, controller register signals, and environmental monitoring data in the multimodal data are aligned according to their respective timestamps. Interpolation methods are used to fill in missing time points, such as linear interpolation or spline interpolation. Outlier removal is performed on the aligned multimodal data sequence, with outlier identification based on statistical methods. Noise suppression algorithms are applied to the outlier-removed data, such as moving average filtering or Kalman filtering. After processing, a cleaned, high-quality dataset is generated.

[0057] S1.3. The cleaned high-quality dataset is uniformly resampled according to modality type and sampling frequency, and the uniform high-quality dataset is normalized and standardized by unit to generate a standardized numerical matrix.

[0058] Specifically, the voltage waveform data, current waveform data, controller register signal, and environmental monitoring data in the cleaned high-quality dataset are uniformly resampled according to their corresponding sampling frequencies. The sampling frequency is derived from the parameter configuration of the acquisition device; for example, the sampling frequency for voltage and current waveform data is 5000 Hz, the sampling frequency for controller register signal is 1000 Hz, and the sampling frequency for environmental monitoring data is 1 Hz. Resampling algorithms, including interpolation or downsampling methods, are used to bring the multimodal data to a uniform sampling frequency, forming a uniformly resampled high-quality dataset. Subsequently, normalization processing is performed on the uniformly resampled high-quality dataset, using the max-min normalization method to map it to the [0,1] interval. Then, unit standardization processing is performed on the normalized data, using the Z-Score standardization method with a mean of 0 and a standard deviation of 1. Finally, a standardized numerical matrix is ​​generated.

[0059] S1.4 Organize the standardized numerical matrix into a three-dimensional structure according to time steps, spatial nodes and modal dimensions to generate a standardized spatiotemporal data cube.

[0060] Specifically, the standardized numerical matrix is ​​divided into time step dimension, spatial node dimension, and modal dimension. The time step dimension corresponds to the sampling time series after unified resampling, the spatial node dimension corresponds to the physical location number of the acquisition point, and the modal dimension corresponds to four modes: voltage waveform data, current waveform data, controller register signal, and environmental monitoring data. All standardized values ​​on the spatial nodes of the time step are arranged sequentially to construct a three-dimensional data structure with the shape of "time step number × spatial node number × modal number", thus completing the generation of the standardized spatiotemporal data cube.

[0061] S2. Use the causal discovery algorithm to mine causal associations and identify dependency paths in the standardized spatiotemporal data cube, construct a causal topology graph, and use a spatiotemporal graph attention network to perform cross-modal feature fusion to generate a high-order feature tensor.

[0062] S2.1. Expand the standardized spatiotemporal data cube into multivariate time series samples, and use conditional independence test and time delay analysis to identify potential causal relationships between state variables in the multivariate time series samples.

[0063] Specifically, standardized values ​​of all spatial nodes and modal dimensions corresponding to each time step are extracted sequentially according to the time dimension, and combined to form a multivariate time series sample. The state variables of each time series sample correspond to voltage waveform data, current waveform data, controller register signals, and environmental monitoring data, and the names and order of the state variables are kept consistent. For any two state variables in the multivariate time series sample, based on the existing conditional independence test method, the statistical independence of the two state variables is calculated under the condition of controlling the other state variables, and it is determined whether there is a direct dependency relationship between the two state variables. For each pair of state variables, the time delay analysis method is used to calculate the time delay parameter between the two state variables. Specifically, the delay correlation analysis is performed on the multivariate time series sample through a sliding time window to determine the delay duration. Combining the dependency judgment of the conditional independence test and the delay duration of the time delay analysis, potential causal relationship pairs between state variables in the multivariate time series sample are identified.

[0064] It should also be noted that the specific steps for determining whether there is a direct dependency between two state variables are as follows: Identify the two state variables to be tested and the set of other state variables under control conditions; select an appropriate conditional independence test statistic, such as a statistic based on partial correlation coefficients or mutual information, and calculate the conditional dependency between the two state variables under the control of the other state variables; calculate the corresponding p-value according to a preset significance level, such as 0.05 in this example; compare the calculated p-value with the significance level. If the p-value is less than the significance level, it is determined that there is a direct dependency between the two state variables; otherwise, it is determined that the two state variables are conditionally independent and there is no direct dependency.

[0065] S2.2 Construct the identified potential causal relationships into directed edges, and use all state variables as nodes to generate a preliminary causal graph structure.

[0066] Specifically, for each pair of state variables in the identified potential causal relationships, a directed edge is constructed according to the causal direction. First, all state variables are used as nodes, and the identified potential causal relationships are traversed one by one. For the starting and ending state variables of each potential causal relationship, a directed edge is established from the starting state variable to the ending state variable. Finally, all directed edges and the set of state variable nodes are combined to form a preliminary causal graph structure.

[0067] S2.3. Apply a structural optimization algorithm to the initial causal graph structure to eliminate false associations and redundant paths, and obtain a sparse causal topology graph.

[0068] Specifically, for the initial causal graph structure, a significance test is performed on each directed edge. Statistical significance is used to determine whether a directed edge is a spurious association, and directed edges that do not meet the significance condition are removed. Based on path cover analysis, multiple paths between all node pairs are traversed to identify directed edges containing redundant paths. Specifically, if a directed edge on one path exists on another path without affecting causal connectivity, the directed edge is identified as a redundant path and removed. After significance testing and path cover analysis, and the removal of spurious and redundant directed edges, a sparsified causal topology graph is generated. Spurious associations are identified based on the significance level in the conditional independence test results, and redundant paths are identified based on the path cover principle in graph theory.

[0069] S2.4. Based on the node dependencies and temporal order in the sparse causal topology graph, label the directionality and strength of the edges to form a causal topology graph.

[0070] Specifically, based on the dependencies between nodes in the sparse causal topology graph, the starting and ending node pairs connected by all directed edges are first extracted as the basis for causal relationships. Based on the timestamp information of the multivariate time series samples, the state variable values ​​of each node are sorted in ascending order according to time sequence to analyze the temporal relationship between nodes and confirm that the direction of the directed edges is consistent with the time flow. Next, the strength of each directed edge is calculated, using the statistic from the conditional independence test or the correlation coefficient from time delay analysis as the strength index; a specific example is using the F-statistic value of the Granger causality test. Finally, the directed edges labeled with direction and strength are combined with all state variable nodes to construct the causal topology graph. Node dependencies originate from conditional independence tests and the identification of potential causal associations.

[0071] It should also be noted that the specific process of sorting the state variable values ​​of each node in ascending order according to time sequence includes: obtaining the timestamp data corresponding to all state variables in the multivariate time series sample, ensuring that the timestamp format is uniform and accurate; arranging the state variable data of each node in ascending order according to the timestamp from smallest to largest to form a time series; using the same time axis reference for all nodes, mapping the time series of state variables of each node onto the time axis to ensure time step synchronization; and outputting the sequence of state variable values ​​of each node arranged in chronological order as the basis for subsequent time series analysis.

[0072] S2.5. Encode the causal topological graph into a graph structure tensor, construct a graph representation processing structure, use the attention mechanism to fuse the association information of nodes in the causal topological graph, extract multi-scale representations that reflect the structural hierarchical features, and generate a unified modal representation.

[0073] Specifically, the process of encoding a causal topological graph into a graph structure tensor is as follows: based on all state variables in the causal topological graph as nodes, the feature vectors of each node are extracted and arranged into a node feature matrix; based on the directed edges marked in the causal topological graph, the connection relationships of the edges are converted into an adjacency matrix, and the elements of the adjacency matrix represent the existence and weight of the edges between the corresponding nodes; the node feature matrix and the adjacency matrix together form a graph structure tensor, which is used as the input for subsequent graph representation processing.

[0074] The specific steps for constructing the graph representation processing structure are as follows: The graph attention mechanism in graph neural networks is used to process the graph structure tensor input. Specifically, the attention weights between each node and its neighboring nodes are calculated. These weights are obtained by linear transformation of the node features and the dot product between nodes, and then normalized. The weighted neighbor node features are aggregated to form the attention aggregated feature representation of each node. Multi-layer attention calculations are repeated to achieve multi-scale feature extraction. The aggregated features of each node are then fused to generate a unified modality representation.

[0075] S2.6. Using the time attention mechanism, the uniform modality representation is weighted and aggregated within the historical time window to obtain the time evolution feature representation.

[0076] Specifically, the process involves extracting a sequence of feature vectors from multiple time steps within a historical time window from the unified modality representation, for example, extracting the unified modality representation of the 10 most recent consecutive time steps. Next, a linear transformation is performed on the feature vector of each time step to obtain the query vector, key vector, and value vector. Then, the dot product similarity between the query vector and the key vector is calculated, scaled, and normalized using the Softmax function to obtain the attention weight coefficient for each time step. The value vector of each time step is then multiplied by the corresponding attention weight coefficient and summed using a weighted sum to obtain a time-weighted aggregated representation. Finally, the weighted aggregated result is output as the time-evolution feature representation.

[0077] S2.7. The unified modal representation and the temporal evolution feature representation are fused through feature concatenation to generate a high-order feature tensor.

[0078] Specifically, the process involves obtaining the feature vectors of the unified modality representation and the temporal evolution feature representation, for example, where both are vector sequences of the same length. Next, the feature vectors of the unified modality representation are arranged sequentially to form the first tensor dimension, and the feature vectors of the temporal evolution feature representation are arranged in the same order to form the second tensor dimension. Following a feature concatenation method, the feature vectors of the unified modality representation and the temporal evolution feature representation are concatenated along the feature dimensions, where concatenation involves appending each element of the unified modality representation sequentially to the corresponding element of the temporal evolution feature representation. The resulting high-order feature tensor contains all the feature information from both the unified modality representation and the temporal evolution feature representation, serving as input for subsequent processing.

[0079] It should also be noted that the specific steps for obtaining the feature vectors of the unified modality representation and the temporal evolution feature representation are as follows: For the causal topological graph, after graph structure tensor construction and attention mechanism processing, the unified modality representation corresponding to each node is obtained. The unified modality representation is the numerical vector of the node on the multi-scale structural hierarchical features. Next, the unified modality representation is weighted and aggregated within the historical time window using the temporal attention mechanism to obtain the temporal evolution feature representation. The temporal evolution feature representation is the feature vector after fusing temporal information. The unified modality representation and the temporal evolution feature representation are output as two sets of feature vectors for subsequent feature fusion processing.

[0080] S3. Divide the high-order feature tensors into meta-learning task pools according to different machine models and environmental conditions, and use a two-layer optimization strategy to train a cross-machine model general parameterization framework.

[0081] S3.1 Classify the high-order feature tensors according to the device type and environmental conditions, divide them into multiple meta-learning task pools, and initialize the global parameters of the cross-device general parameterization framework based on the samples in the meta-learning task pools.

[0082] Specifically, high-order feature tensors are classified based on aircraft type and environmental conditions. First, the aircraft type and environmental conditions are obtained, where the aircraft type comes from the basic information record of the equipment, and the environmental conditions come from parameters such as temperature, humidity, and air pressure collected by sensors. Then, the high-order feature tensors are grouped according to the aircraft type and environmental conditions to form multiple meta-learning task pools. The samples in the meta-learning task pools are extracted from historical operating data under the corresponding aircraft type and environmental conditions using preprocessed high-order feature tensors. The cross-aircraft general parameterization framework includes a neural network structure with an input layer, multiple hidden layers, and an output layer, as well as weight matrices and bias vectors for parameter sharing and adjustment. Based on the high-order feature tensor samples in each meta-learning task pool, the global parameters of the cross-aircraft general parameterization framework are initialized using gradient aggregation or parameter averaging methods to improve the generalization ability of the global parameters.

[0083] S3.2 In each meta-learning task pool, local task optimization is performed on the cross-machine general parameterization framework using task-specific data to obtain task adaptation parameters.

[0084] Specifically, in each meta-learning task pool, task-specific data is first extracted from the operation logs and sensor monitoring data under the corresponding device type and environmental conditions. Then, using the global parameters of the cross-device general parameterization framework as initialization parameters, multiple rounds of gradient descent optimization iterations are performed based on the task-specific data. Each round includes forward propagation to calculate the loss function and backpropagation to update the initialization parameters. The optimization stops by limiting the number of training rounds or the convergence criterion of the loss function, thus obtaining the task adaptation parameters and realizing the adjustment and adaptation of local task parameters.

[0085] S3.3 Aggregate the task adaptation parameters from all meta-learning task pools, and use a two-layer optimization strategy to update the initialized global parameters to obtain the optimized global parameters.

[0086] Specifically, the task adaptation parameters obtained through local task optimization in all meta-learning task pools are collected; then, the initialized global parameters are used as the outer optimization variables, and the task adaptation parameters in each task pool are used as the inner optimization results to calculate the loss function of each task pool; by calculating the weighted average of the loss functions of all task pools, the outer gradient descent is performed to update the global parameters, completing one iteration of global parameter optimization; the inner local task optimization and outer global parameter update are repeated until the preset number of iterations or the loss function convergence condition is met, and finally the optimized global parameters are obtained.

[0087] It should also be noted that the preset iteration count step includes: determining the maximum number of iterations based on actual needs and computing resources, for example, setting it to one hundred times; setting a convergence threshold for the loss function, for example, determining convergence when the decrease in the loss function over five consecutive iterations is less than 0.1%; the step of determining the convergence condition of the loss function includes: calculating the current loss function value after each iteration and comparing it with the loss function value of the previous iteration, calculating the difference between the current loss function value and the loss function value of the previous iteration; if the difference is less than the preset convergence threshold and the preset number of iterations is met consecutively, the loss function is considered to have converged, and the iteration stops; if the maximum number of iterations is reached but the convergence condition is still not met, the iteration stops, and the current difference is taken as the final result.

[0088] S3.4 Repeat the iterative process of local task optimization and global parameter update until the cross-machine general parameterization framework training converges, and output the trained cross-machine general parameterization framework.

[0089] Specifically, when training the cross-device general parameterization framework, in each iteration, local task optimization is first performed on the global parameters of the cross-device general parameterization framework for the current iteration using the corresponding task-specific data for each meta-learning task pool, obtaining multiple task-adaptive parameters. Then, all task-adaptive parameters are aggregated, and a two-layer optimization strategy is used to update the global parameters of the cross-device general parameterization framework for the current iteration. After completing one round of local task optimization and global parameter update, it is determined whether the preset number of iterations or the convergence condition of the loss function are met. If not, the next round of local task optimization and global parameter update operation is continued until the stopping condition is met. Finally, the cross-device general parameterization framework corresponding to the current iteration is output as the trained cross-device general parameterization framework.

[0090] It should also be noted that when determining whether the preset number of iterations or the convergence condition of the loss function are met, the maximum number of iterations threshold and the loss function convergence threshold are first set. For example, the maximum number of iterations is one hundred, and the loss function convergence threshold is 0.001. After completing the local task optimization and global parameter update in each round, the current round number is recorded and the global loss function value of the current round is calculated. It is then determined whether the current round number is greater than or equal to the maximum number of iterations threshold. If so, the iteration is terminated. Otherwise, the difference between the global loss function value of the current round and the global loss function value of the previous round is calculated. If the difference is less than or equal to the loss function convergence threshold, convergence is determined, and the iteration is terminated. If none of the conditions are met, the next round of local task optimization and global parameter update operation is performed.

[0091] S4. Based on the trained cross-model general parameterized framework, a composite operating condition generator is constructed, and constraints are applied through a causal regularization loss function to form an extended test condition set.

[0092] S4.1 Based on the trained cross-model general parameterization framework, structural parameters and representation vectors are extracted from the high-order feature tensor and input into the composite operating condition generator for initialization processing to obtain the initial structural parameters of the composite operating condition generator.

[0093] Specifically, based on the trained cross-machine universal parameterization framework, the structural parameters and representation vectors corresponding to the current input sample are extracted from the high-order feature tensor through predefined structural parameter extraction functions and representation vector extraction functions within the framework. The extracted structural parameters and representation vectors are concatenated to form a joint representation vector, which is then input into the composite load condition generator for initialization. Inside the composite load condition generator, the joint representation vector is first linearly transformed to align dimensions, then normalized to eliminate dimensional differences, and initial structural parameters are generated according to initialization rules (e.g., Gaussian initialization with a mean of zero and a standard deviation of one). Finally, the processed vector is output as the initial structural parameters of the composite load condition generator.

[0094] It should also be noted that the specific steps for predefining the structural parameter extraction function and the representation vector extraction function are as follows: Based on the dimensional structure of the high-order feature tensor, the dimensions used to describe the changes in aircraft type and the dimensions used to represent state sequences or time series are jointly classified as structurally relevant dimensions, and the dimensions used to represent feature semantic information are classified as semantically relevant dimensions; a structural parameter extraction function is constructed based on the structurally relevant dimensions, specifically by performing multiple one-dimensional convolution operations on the structurally relevant dimensions, with each convolution operation corresponding to a sliding window of a fixed size, and extracting the corresponding convolution features as structural parameters; a representation vector extraction function is constructed based on the semantically relevant dimensions, specifically by expanding the semantically relevant dimensions into a vector sequence and inputting it into a multi-head attention encoding network, capturing the dependencies between different semantics through the attention mechanism, and outputting a vector of a unified dimension as the representation vector; labeled high-order feature tensor samples containing different aircraft types and different environmental conditions are used as training data to perform supervised training on the structural parameter extraction function and the representation vector extraction function respectively; the network parameters of the trained structural parameter extraction function and the representation vector extraction function are solidified and retained as fixed structures in the trained cross-aircraft general parameterization framework.

[0095] S4.2. Use real working condition samples to preheat and train the initial structural parameters of the composite working condition generator. After training, a composite working condition generator with preliminary working condition construction capability is obtained, generating preliminary working condition samples. Based on the causal relationship between nodes in the causal topology graph, a causal regularization loss function is constructed.

[0096] Specifically, real-world operating condition samples are constructed based on high-frequency monitoring data recorded in historical operation databases of various aircraft models. These samples include input feature tensors and corresponding output tensors, covering typical operating phases such as startup, steady state, load variation, and fault conditions. Keeping the initial structural parameters in the composite operating condition generator unchanged, the structural parameters and representation vectors extracted from the trained cross-model general parameterization framework are input into the composite operating condition generator along with the real-world operating condition samples. Preheating training is performed using mean squared error as the loss function until the loss value decreases to a stable range, for example, below one percent. After training, the trained composite operating condition generator is used to perform forward generation operations on the higher-order feature tensors. A preliminary operating condition sample is generated. Based on the analysis results of the temporal structure and operating mechanism of each variable in the real operating condition sample, a causal topology graph is constructed. In the causal topology graph, nodes correspond to variable names, and edges represent explicit causal paths. Causal paths are determined by Granger causality analysis combined with engineering logic. For example, the causal order between fuel supply rate, combustion chamber temperature and exhaust emission concentration is confirmed based on physical mechanisms and hysteresis correlation. For each pair of variables represented by a directed edge in the causal topology graph, a causal consistency loss term is constructed. The degree of deviation between the generated results of variables is used as a regularization term to form a causal regularization loss function, which is then weighted and combined with the main loss function to form a total loss function for subsequent training.

[0097] S4.3. Based on the preliminary working condition samples and the constructed causal regularization loss function, the generation loss and the causal regularization loss based on the causal consistency metric are calculated using the adversarial training method. The parameters of the composite working condition generator are jointly optimized to obtain the optimized composite working condition generator and output the working condition samples.

[0098] Specifically, an adversarial training structure is constructed, comprising a generator and a discriminator. A composite working condition generator acts as the generator, using initial working condition samples as input to generate pseudo-working condition samples. Simultaneously, both real working condition samples and the generated pseudo-working condition samples are input into the discriminator, which employs a binary classification structure and outputs true / false labels. The adversarial loss generated by the composite working condition generator for pseudo-working condition samples is used as the primary optimization objective. The adversarial loss is defined based on the discriminator's misclassification rate; for example, the cross-entropy loss function is used to calculate the probability error of a pseudo-sample being classified as real. Based on the causal regularization loss function constructed in the previous step, the probability error of the pseudo-working condition samples being classified as real is calculated. The causal consistency deviation among variables in the causal topology graph is used to obtain the causal regularization loss. The adversarial loss and the causal regularization loss are weighted and summed to form a joint loss function. The weight parameters of the adversarial loss and the causal regularization loss can be set with empirical values, such as 0.8 for the adversarial loss and 0.2 for the causal regularization loss. The parameters of the composite working condition generator are updated using gradients using the joint loss function. Existing optimization algorithms, such as the Adam optimization algorithm, are used for multiple rounds of iterative updates until the loss converges. The jointly optimized composite working condition generator is output, and the working condition samples are finally output based on the high-order feature tensor.

[0099] It should also be noted that the specific steps for constructing the adversarial training structure containing a generator and a discriminator are as follows: A composite working condition generator is used as the generation end, with high-order feature tensors input into the composite working condition generator, and pseudo-working condition samples are output; a discriminator is constructed using a multi-layer fully connected neural network structure, with the input dimension consistent with the pseudo-working condition samples, and the output layer being a single neuron structure, outputting continuous values ​​between zero and one, representing the probability that the input sample is a real working condition sample; the real working condition samples and the pseudo-working condition samples output by the composite working condition generator are concatenated to form a training sample set, which is input into the discriminator and supervised learning is performed using binary classification labels, with real working condition samples labeled as one and pseudo-working condition samples labeled as zero, and the discrimination error is calculated using the cross-entropy loss function; the discriminator parameters are frozen, and backpropagation optimization is performed only on the composite working condition generator to increase the probability that pseudo-working condition samples are judged as real working condition samples, while the causal regularization loss and the cross-entropy adversarial loss are weighted and summed to form a joint loss function, which is used to update the composite working condition generator parameters.

[0100] The weight parameters of adversarial loss and causal regularization loss can be set empirically. Specifically, the initial weight values ​​are set before training, for example, the adversarial loss weight is set to 0.8 and the causal regularization loss weight is set to 0.2, which are used as coefficients in the joint loss function. The ratio can be adjusted by observing the quality of generated samples and causal consistency indicators in multiple training batches. If the probability of pseudo-work case samples being identified as real samples is too low, the weight of adversarial loss is appropriately increased. If the causal relationship of pseudo-work case samples deviates significantly in the causal topology graph, the weight of causal regularization loss is appropriately increased until the gradient magnitudes of the two types of losses reach a balance during training.

[0101] S4.4 Perform causal consistency testing and distribution characteristic screening on the output working condition samples to form an extended test condition set.

[0102] Specifically, based on the constructed causal topology graph, the causal consistency measure among variables in the output working condition sample is calculated, and samples that meet the preset causal consistency threshold are selected as qualified causal consistency samples. The statistical distribution characteristics of the output working condition sample are extracted, including mean, variance, skewness, and kurtosis. Multidimensional statistical test methods are used to compare the distribution characteristics of the real working condition sample, and abnormal samples with significant deviations are removed. Finally, the sample sets that are qualified for causal consistency and whose distribution characteristics meet the requirements are merged to form an extended test working condition set.

[0103] It should also be noted that the specific steps for selecting samples that meet the preset causal consistency threshold as qualified causal consistency samples are as follows: For each output working condition sample, calculate the causal consistency measure value corresponding to each causal relationship in the causal topology diagram. Commonly used measures include the causal effect strength or conditional independence test index; compare the causal consistency measure value of each causal relationship with the preset causal consistency threshold one by one. The preset causal consistency threshold is 0.7, which means that the relationship with a causal consistency measure value of not less than 0.7 is considered to meet causal consistency; only the working condition samples with all key causal relationship causal consistency measure values ​​not less than the preset causal consistency threshold are retained as qualified causal consistency samples.

[0104] S5. A multimodal deep reinforcement learning algorithm is used to optimize the control parameters of the extended test case set to obtain optimized control parameters.

[0105] S5.1. Using an extended test case set, construct a multimodal environment state representation, and use the policy gradient method to evaluate the policy of the multimodal environment state representation and generate control parameter policies.

[0106] Specifically, based on the extended test case set, multiple modal features, including temporal features, spatial features, and statistical features, are first extracted and fused to form a multimodal environment state representation. A policy network is then constructed, taking the multimodal environment state representation as input and outputting the probability distribution of the control parameter policy. Using the policy gradient method, the corresponding actions are executed by sampling the control parameter policy, and the payoff function value corresponding to the action is calculated. The gradient of the policy network parameters is estimated and updated based on the payoff function. The sampling, execution, and policy network parameter update process is repeated until the policy network parameters converge, and the final control parameter policy is output.

[0107] It should also be noted that the specific process of extracting multimodal features and constructing the policy network is as follows: For each working condition sample in the extended test case set, temporal features are extracted, including the change sequence features of each working condition variable in the working condition sample in the time dimension. A one-dimensional convolutional neural network is used to extract local time patterns and output a temporal feature tensor. Spatial features are extracted, including the joint relationship between different working condition variables at the same time step. A two-dimensional convolutional neural network is used to extract spatial correlation features in the variable dimension and output a spatial feature tensor. Statistical features are extracted, including indicators such as mean, variance, skewness, and kurtosis. These are directly calculated using statistical calculation methods and output as a statistical feature vector. The temporal features, spatial features, and statistical features are concatenated and fused to form a multimodal environmental state representation.

[0108] A policy network is constructed with the input dimension consistent with the state representation of the multimodal environment. The network structure adopts a multi-layer fully connected neural network with ReLU activation function in the hidden layer and Softmax function in the output layer to generate the probability distribution of the control parameter policy. The output is used to represent the sampling probability of each control parameter policy under the current state representation of the multimodal environment.

[0109] S5.2 Input the control parameter strategy into the simulation environment corresponding to the extended test condition set, monitor the simulation response process, collect control effect data, and calculate feedback rewards based on preset performance indicators.

[0110] Specifically, the control parameter strategy generates specific control parameter values ​​according to the sampling method, and these values ​​are input into the simulation environment corresponding to each set of operating condition samples in the extended test case set. The simulation environment is a reproducible process response simulation environment established on an industrial process simulation platform based on the initial operating condition variables recorded in the operating condition samples. The operating condition variables include control inputs, disturbances, and controlled variables. Then, during the simulation environment operation, the response process of key controlled variables is monitored in real time, and the output time series is collected as control effect data. Next, each set of simulation response data is processed according to preset performance indicators to calculate the corresponding feedback reward value. The feedback reward is used to guide subsequent strategy updates.

[0111] It should also be noted that the specific steps for setting performance indicators are as follows: Determine the performance dimensions, including response speed, steady-state error, overshoot, and settling time; construct evaluation formulas for each performance dimension, where response speed calculates the shortest time for the controlled variable to reach the target value and stabilize; steady-state error calculates the average deviation during the steady-state phase; overshoot calculates the percentage of the maximum deviation exceeding the target value; and settling time calculates the time point at which the variable first enters the target range and remains stable; set target values ​​for each performance dimension, such as response speed ≤ 10 seconds, steady-state error ≤ 1%, overshoot ≤ 5%, and settling time ≤ 12 seconds; assign weight values ​​to each performance dimension and integrate the evaluation results in a weighted manner to form a comprehensive feedback reward, for example, response speed weight is 0.4, steady-state error weight is 0.3, overshoot weight is 0.2, and settling time weight is 0.1.

[0112] S5.3 Based on feedback rewards, the control parameter strategy is adjusted using an optimization algorithm. The training is repeated until the preset performance index is reached, and the optimized control parameters are obtained.

[0113] Specifically, based on the feedback reward corresponding to the control parameter strategy, and using the feedback reward as the evaluation criterion for the objective function, the probability distribution of the control parameter strategy is calculated using the strategy gradient method. The values ​​of each strategy variable in the control parameter strategy are adjusted according to the gradient direction to update the control parameter strategy. Then, the updated control parameter strategy is input again into the simulation environment corresponding to the extended test case set, and control effect data, feedback rewards, and strategy adjustment operations are repeatedly collected. This iterative training continues to optimize the control parameter strategy until the preset performance indicators corresponding to the feedback rewards all meet the set target standard values, such as response speed no greater than 10 seconds, steady-state error no greater than 1%, overshoot no greater than 5%, and settling time no greater than 12 seconds. Finally, optimized control parameters that meet the preset performance indicator requirements are output.

[0114] S6. Construct a digital twin virtual operating condition instance using optimized control parameters, perform dynamic time warping verification, and generate a standardized verification report.

[0115] S6.1. By expanding the test condition set and optimizing the control parameters, a digital twin virtual operating condition instance is constructed.

[0116] Specifically, multiple typical operating condition samples from the extended test condition set are selected as virtual operating condition input data. Each typical operating condition sample is paired with its corresponding optimized control parameters and input into the simulation environment corresponding to the extended test condition set for multiple rounds of simulation. During the simulation, the state variables, control output values, and environmental response data at each moment are recorded in real time, and the dynamic change characteristics of the control process are collected synchronously. Then, the simulation results are labeled and bound with the operating condition sample data and optimized control parameters in the extended test condition set to establish a one-to-one correspondence between operating condition input, control parameters, and response results. Finally, a complete state evolution sequence and control execution trajectory are generated for each set of typical operating condition samples, forming a digital twin virtual operating condition operation instance set covering different environmental disturbance conditions.

[0117] S6.2. Based on the optimized control parameter input, a digital twin virtual operating condition instance is used to collect the time-series response data of the operating instance and perform dynamic time warping processing with the multimodal data to obtain aligned time-series data.

[0118] Specifically, optimized control parameters are input into the simulation environment corresponding to the extended test case set, and the corresponding digital twin virtual operating case instance is run. Throughout the simulation, time-series response data such as voltage, current, active power, frequency, and load changes are collected. Multimodal data corresponding to the digital twin virtual operating case instance is acquired, including a multimodal environmental state representation composed of temporal, spatial, and statistical features. Feature sequences under each mode are extracted using each time point as an index. Then, a dynamic time warping method is used to align the time-series response data with the time-series data in the multimodal environmental state representation. The optimal alignment path is calculated based on the minimum distance matching principle, and the aligned time-series dataset is output. Finally, based on the aligned time-series data, a dataset sample is established for subsequent causal consistency evaluation.

[0119] S6.3 Based on the aligned time series data, using root mean square error and dynamic time warping distance, calculate the error index and dynamic performance deviation between the digital twin virtual operating condition instance and the extended test condition set, evaluate the response consistency of the digital twin virtual operating condition instance, and generate a standardized verification report.

[0120] Specifically, aligned time-series data from the digital twin virtual operating condition instance and the corresponding real operating conditions in the extended test condition set are extracted to construct matching data pairs. The root mean square error (RMSE) is calculated for each matching data pair, specifically by squaring the difference between the two time-series values ​​at each time point, averaging the differences, and then taking the square root to obtain the RMSE result. Furthermore, the dynamic time warping method is used to calculate the dynamic time warping distance between the digital twin virtual operating condition instance and the extended test condition set, obtaining the shortest path and total distance value for the full sequence matching. Then, the RMSE and dynamic time warping distance are used as error indicators and dynamic performance deviation measurements to evaluate the response consistency of the digital twin virtual operating condition instance. Finally, all error indicator results, dynamic time warping distance values, and consistency judgment conclusions are summarized and output in a unified format to generate a standardized verification report.

[0121] It should also be noted that the specific steps for constructing matching data pairs include: selecting sensor observation time series data corresponding to each set of operating conditions in the extended test condition set as reference data, and clarifying the time index and variable identifier; secondly, selecting the time series response data corresponding to the optimized control parameters generated by the digital twin virtual operating condition instance under the same operating condition input conditions, and extracting data with the same time index and variable identifier as the extended test condition set; for each set of time series data in the extended test condition set, matching the timestamp with the variable identifier one-to-one, and constructing a data pair consisting of "extended test condition set reference data - digital twin virtual operating condition instance response data"; finally, uniformly numbering and structuring all constructed matching data pairs, serving as input data for subsequent calculation of root mean square error and dynamic time warping distance.

[0122] This embodiment also provides a computer device applicable to the electromechanical transient modeling method of multimodal wind turbine based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the electromechanical transient modeling method of multimodal wind turbine based on artificial intelligence as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the electromechanical transient modeling method for multimodal wind turbines based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] In summary, this invention achieves in-depth analysis of potential causal relationships within multimodal spatiotemporal data by employing a causal discovery algorithm to mine causal associations and identify dependency paths in a standardized spatiotemporal data cube. This helps reveal the true causal dependency structure between state variables, avoiding the problems of causal confusion and redundant information interference in traditional correlation analysis. By constructing a causal topology graph and utilizing a spatiotemporal graph attention network to weightedly fuse the association information of nodes in the causal structure, efficient integration of cross-modal features and generation of high-order feature tensors are achieved. Through accurate causal modeling and cross-modal fusion, a solid foundation is laid for achieving highly robust and generalizable electromechanical transient modeling, effectively improving accuracy and adaptability.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence based multi-modal wind turbine electromechanical transient modeling method, characterized in that: The application relates to a wind turbine fault prediction method based on multi-modal data and a wind turbine fault prediction device. Multi-modal data of a wind turbine is collected and preprocessed to generate a standardized space-time data cube; A causal discovery algorithm is used to mine causal correlations and identify dependency paths of the standardized space-time data cube, a causal topology graph is constructed, cross-modal feature fusion is performed through a space-time graph attention network, and a high-order feature tensor is generated; The high-order feature tensor is divided into a meta-learning task pool according to different machine types and environmental conditions, a double-layer optimization strategy is used to train a cross-machine universal parameterized framework, and the specific steps are as follows, The high-order feature tensor is classified according to machine types and environmental conditions, and multiple meta-learning task pools are divided, global parameters of the cross-machine universal parameterized framework are initialized according to samples in the meta-learning task pools; In each meta-learning task pool, the cross-machine universal parameterized framework is locally optimized using task-specific data to obtain task-adaptive parameters; All task-adaptive parameters in the meta-learning task pools are aggregated, and the initialized global parameters are updated using a double-layer optimization strategy to obtain optimized global parameters; The iteration process of local task optimization and global parameter updating is repeated until the cross-machine universal parameterized framework converges, and the trained cross-machine universal parameterized framework is output. Based on the trained cross-machine universal parameterized framework, a composite working condition generator is constructed, and a causal regularization loss function is used for constraint to form an extended test working condition set; A multi-modal deep reinforcement learning algorithm is used to optimize the control parameters of the extended test working condition set to obtain optimized control parameters; The optimized control parameters are used to construct a digital twin virtual working condition running instance, dynamic time warping verification is performed, and a standardized verification report is generated.

2. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 1, wherein: The standardized space-time data cube is generated, and the specific steps are as follows, Voltage waveforms, current waveforms, controller register signals and environmental monitoring data are collected to obtain multi-modal data; The multi-modal data is aligned according to timestamps, and the time-series aligned multi-modal data sequence is subjected to missing value filling, abnormal value elimination and noise suppression treatment to obtain a cleaned high-quality data set; The cleaned high-quality data set is uniformly resampled according to the modal type and sampling frequency, and the unified high-quality data set is subjected to normalization and unit standardization treatment to generate a standardized numerical matrix; The standardized numerical matrix is organized into a three-dimensional structure according to time steps, spatial nodes and modal dimensions to generate a standardized space-time data cube.

3. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 2, wherein: The standardized space-time data cube is expanded into a multivariate time series sample, and condition independence testing and time delay analysis are used to identify potential causal correlation pairs between state variables of the multivariate time series sample; The identified potential causal correlation pairs are constructed into directed edges, and all state variables are used as nodes to generate a preliminary causal graph structure; A structure optimization algorithm is used on the preliminary causal graph structure to eliminate false correlations and redundant paths to obtain a sparse causal topology graph; According to the node dependency relationship and time sequence order in the sparse causal topology graph, the directionality and intensity of the edges are labeled to form a causal topology graph. ​ 4. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 1, wherein: The generating high-order feature tensor comprises the following specific steps, The causal topology graph is encoded into a graph structure tensor, a graph representation processing structure is constructed, the associated information of nodes in the causal topology graph is fused by using an attention mechanism, a multi-scale representation reflecting structural hierarchical features is extracted, and a unified modal representation is generated; The unified modal representation and the time evolution feature representation are fused by using a feature splicing method to generate a high-order feature tensor. The forming of the extended test working condition set comprises the following specific steps, 5. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 1, wherein: Based on the trained cross-model universal parameterization framework, structural parameters and representation vectors are extracted from the high-order feature tensor and input into the composite working condition generator for initialization processing to obtain initial structural parameters of the composite working condition generator; The initial structural parameters of the composite working condition generator are preheated and trained using real working condition samples, and after training, a composite working condition generator with preliminary working condition construction capability is obtained, preliminary working condition samples are generated, and a causal regularization loss function is constructed based on the causal relationship between nodes in the causal topology graph; Based on the preliminary working condition samples and the constructed causal regularization loss function, an optimized composite working condition generator is obtained by jointly optimizing the composite working condition generator parameters using an adversarial training method to calculate a generation loss and a causal regularization loss based on a causal consistency measure, and outputting working condition samples; The output working condition samples are subjected to causal consistency inspection and distribution feature screening to form an extended test working condition set. The obtaining of the optimized control parameter comprises the following specific steps, 6. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 1, wherein: The extended test working condition set is used to construct a multi-modal environment state representation, and a strategy gradient method is used to evaluate the multi-modal environment state representation to generate a control parameter strategy; The control parameter strategy is input into the simulation environment corresponding to the extended test working condition set, the simulation response process is monitored, control effect data is collected, and feedback rewards are calculated according to preset performance indicators; According to the feedback rewards, an optimization algorithm is used to adjust the control parameter strategy, and the training is repeated until the preset performance indicators are reached to obtain the optimized control parameter. The generating of the standardized verification report comprises the following specific steps, 7. The artificial intelligence based multi-modal wind turbine electro-mechanical transient modeling method of claim 1, wherein: A digital twin virtual working condition running instance is constructed by using the extended test working condition set and the optimized control parameter; Based on the optimized control parameter, time series response data of the digital twin virtual working condition running instance are collected and dynamically time warping processed with the multi-modal data to obtain aligned time series data; Based on the aligned time series data, root mean square error and dynamic time warping distance are used to calculate error indicators and dynamic performance deviations between the digital twin virtual working condition running instance and the extended test working condition set, and the response consistency of the digital twin virtual working condition running instance is evaluated to generate a standardized verification report. The processor executes the computer program to realize the steps of the artificial intelligence-based multi-modal wind turbine electromechanical transient modeling method of any one of claims 1-7.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the artificial intelligence-based multi-modal wind turbine electromechanical transient modeling method of any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: ​

Images