A Method for Identifying the Operating Status of Offshore Direct-Drive Wind Turbines Based on an Improved One-Dimensional Convolutional Neural Network

Abstract

This invention relates to the field of wind farm operation status identification technology, and more particularly to a method for identifying the operation status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network. The method compresses state labels through feeder iso-modeling, reducing model complexity. An improved one-dimensional convolutional neural network model is constructed, integrating dilated convolution and average pooling to enhance multi-scale feature extraction capabilities, thereby improving the accuracy on both the training and test sets. Global average pooling layers replace fully connected layers, reducing the number of parameters and avoiding overfitting. This effectively solves the gradient problem and insufficient multi-scale feature capture issues inherent in traditional one-dimensional convolutional neural network models when handling complex wind farm operation status identification. The model achieves high accuracy on both the training and test sets, and maintains high accuracy even under different Gaussian noise levels, demonstrating good robustness and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of wind farm operation status recognition technology, and in particular to a method for recognizing the operation status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network. Background Technology

[0002] With the accelerated construction of my country's new power system, wind power is playing an increasingly important role in the power system. Unlike onshore wind power, offshore wind resources are stable, the terrain is simple, and the wind turbines are mainly direct-drive, with similar operating environments and centralized grid connection methods, resulting in higher consistency in their status. Identifying their status information is more conducive to the rapid action of stability control devices. However, the high salt spray and high humidity environment at sea exacerbates equipment corrosion and accelerates the aging of electrical connection points and equipment, leading to more sudden changes in operating status. Coupled with high operation and maintenance costs, there is an urgent need for accurate identification of wind turbine operating status to prevent failures.

[0003] Traditional wind farms consist of dozens of wind turbines, each with multiple states. Traditional identification methods lead to a significant increase in model complexity and a decrease in accuracy. When processing time-series electrical data from wind turbines, traditional one-dimensional convolutional neural networks suffer from limited receptive fields of convolutional kernels, making it difficult to capture long-range fault features. Single-scale convolutions cannot adapt to the significant differences between transient and steady-state faults, and max-pooling operations amplify noise interference, reducing model robustness and resulting in insufficient extraction of sequence features, ultimately affecting the accuracy of state identification.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a patent title that effectively solves the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network includes the following steps: Acquire electrical quantity data of offshore direct-drive wind farms, define the wind farm's operating status, and set status labels; The electrical quantity data and the status label are associated to form a raw dataset, which is then preprocessed. An improved one-dimensional convolutional neural network model was constructed, and the model training parameters were configured for dynamic training optimization. The improved one-dimensional convolutional neural network model is trained using the preprocessed original dataset until the accuracy on the validation set converges. The trained model is deployed to the wind farm monitoring system to process electrical quantity data in real time and output status identification results.

[0007] Furthermore, the acquisition of wind farm electrical quantity data includes the following steps: A wind farm simulation model was built based on the PSASP platform; Based on the wind farm topology, a preset number of wind turbines on the same feeder are equivalent to a single virtual machine group; By simulating power grid faults and disturbances, electrical quantity data of the equivalent feeder are collected.

[0008] Furthermore, in the equivalent settings of the wind farm simulation model, for N wind turbines on the same feeder, the electrical quantity X of a single unit is collected. single and feeder electrical quantity X line Verify the relation X line ≈N*X single After its establishment, feeder data was adopted to replace stand-alone data.

[0009] Furthermore, the original dataset is preprocessed, including data cleaning, normalization, and noise-resistant data augmentation enabled during the training phase. The noise-resistant data augmentation is used to simulate field noise, missing measurements, and electromagnetic interference. The noise reduction data enhancement includes injecting color noise, power frequency and its harmonic interference, Gaussian noise and impulse noise with a signal-to-noise ratio between 20 and 40 dB in different channels; The color noise, power frequency and its harmonic interference, Gaussian noise, and impulse noise are applied synchronously to all channels; the noise-resistant data enhancement is only effective on the training set after the training and validation sets are divided, and is turned off during the online inference phase.

[0010] Furthermore, the construction of the improved one-dimensional convolutional neural network model includes the following steps: The input layer is connected to a Gaussian noise layer, which is activated only during the training phase. The convolutional module includes at least one dilated convolutional layer, introducing a gap controlled by the dilation rate r within the convolutional kernel, satisfying: E eks =K+(K-1)*(r-1), E eks The effective receptive field size for dilated convolution; An adaptive dilation rate scheduling module is introduced, which sets multiple dilated convolutional branches in parallel on the same convolutional layer to capture temporal features at different scales. The dilation rate of the branches is set to a value of [value missing]. The outputs of each branch are fused by channel-gated attention weighting to obtain the output of this layer; During training, the weights of the inflation rate are dynamically adjusted by optimizing the objective function. Specifically, bandwidth scores are calculated for electrical quantity data in each training mini-batch, and the bandwidth scores are integrated into the calculation of channel gating attention. The weights of the dilated convolution branches are adaptively adjusted based on the bandwidth scores. After features are compressed by a global average pooling layer, they are fed into a fully connected layer. The output layer uses the Softmax activation function to generate a 5-dimensional state probability vector.

[0011] Furthermore, when the configuration model training parameters are dynamically trained and optimized, one of the above-mentioned main enhancements is randomly selected in each training batch and a small noise enhancement can be superimposed. The enhancement intensity parameter is adaptively sampled within a preset range to improve the recall rate for rare working conditions. The configuration model training parameters are dynamically trained and optimized, including the following steps: The Adam optimizer was used with an initial learning rate of 0.001, which decayed to 0.1 times the original value every 400 rounds. The model parameters are constrained by L2 regularization, and the L2 regularization coefficient is set to 1×10. -4 ; An early stop mechanism is set up, which calculates the accuracy of the validation set in real time and stops training when the accuracy fluctuation is less than 0.1% for 10 consecutive rounds.

[0012] Furthermore, the operation of the dilated convolutional layer satisfies: The formula for calculating the output signal y is: Where r=2, the convolution kernel is w, the convolution kernel size is K=3, the receptive field is expanded to 5, i is the output signal index, and j is the convolution kernel index; It is combined with standard convolutional layers to form a multi-scale feature extraction chain.

[0013] Furthermore, after the convolution module, average pooling is used to calculate the average value of the local region. ; Where W is the pooling window size and S is the step size.

[0014] Furthermore, the real-time status identification of the wind farm's operating status includes the following steps: The electrical quantity data collected in real time by the SCADA system of the wind farm is cleaned and normalized. The preprocessed electrical quantity data is input into the trained improved one-dimensional convolutional neural network model. The state label corresponding to the maximum value of the Softmax output is taken as the recognition result.

[0015] Furthermore, based on the voltage ride-through standard, the operating status of wind farms is defined as including normal state, high voltage ride-through state, low voltage ride-through state, low voltage ride-through recovery state, and grid disconnection state. When the identification result is that the grid is disconnected, a tripping command is sent to the grid control system; When the identification result is a high voltage ride-through state or a low voltage ride-through state, the reactive power compensation device control signal is activated.

[0016] The technical solution of this invention can achieve the following technical effects: This invention targets offshore direct-drive wind farms. It compresses state labels through feeder-based equivalent modeling, reducing model complexity. An improved one-dimensional convolutional neural network (CNN) model is constructed, integrating dilated convolution and average pooling to enhance multi-scale feature extraction capabilities, thereby improving accuracy on both the training and test sets. Global average pooling layers replace fully connected layers, reducing the number of parameters and preventing overfitting. By introducing noise reduction enhancement, dilated convolution, adaptive dilation rate scheduling, and average pooling techniques, the invention effectively addresses the gradient problem and insufficient multi-scale feature capture issues inherent in traditional one-dimensional CNN models when handling complex state identification at offshore direct-drive wind farms. The improved CNN model achieves accuracy of at least 98.5% on the training and test sets, and 97.5% on the test sets, respectively. Compared to traditional CNN models, the average accuracy is improved by approximately 1.15% and 0.75%, respectively. This model maintains high accuracy under different noise levels, demonstrating good robustness and generalization ability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram illustrating the fault ride-through requirements of wind turbine generators in this invention; Figure 2 This is a schematic diagram illustrating the division of unit operating states in this invention; Figure 3 This is a flowchart of the state recognition algorithm for a one-dimensional convolutional neural network in this invention; Figure 4 This is a structural diagram of the improved one-dimensional convolutional neural network model in this invention; Figure 5 This is a simulation model diagram of a wind farm in this invention; Figure 6 This is a schematic diagram showing the changes in electrical quantities between the feeder and the 11 generating units under fault conditions in this invention; Figure 7 This is a simulation model diagram of the equivalent wind farm in this invention; Figure 8 This is a simulation diagram of the direct-drive fan system in this invention; Figure 9 This is a schematic diagram comparing the accuracy of a single 1D-CNN result with the accuracy of the improved GAP-1DCNN training set in an embodiment of the present invention; Figure 10 This is a schematic diagram comparing the accuracy of a single 1D-CNN result with that of the improved GAP-1DCNN test set in an embodiment of the present invention; Figure 11 This is a schematic diagram showing the results of running 1D-CNN 10 times in an embodiment of the present invention; Figure 12 This is a schematic diagram showing the results of running the improved 1D-CNN 10 times in an embodiment of the present invention; Figure 13 This is a comparison chart of the accuracy of the test set under different Gaussian noise levels in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] When a fault or disturbance occurs in the power system of a wind farm, the direct-drive wind turbine may enter a voltage ride-through state or be directly disconnected from the grid. The change in its operating status will affect the safe and stable operation of the grid. Quickly and accurately identifying the real-time operating status of the unit is conducive to the emergency monitoring device of the system to prevent and control emergencies in a timely manner, and to build a new fault defense system for the power system.

[0022] According to the fault ride-through standard for wind turbines, when a disturbance or fault occurs at the grid connection end of the turbine, its terminal voltage will rise or fall within a certain range. At this time, the turbine should have low voltage or high voltage ride-through capability to maintain continuous operation without disconnecting from the grid.

[0023] See Figure 1As shown, this describes the fault ride-through capability of wind turbines to maintain continuous operation during grid faults. When the voltage amplitude at the turbine's grid connection point is between curves 1 and 2, the turbine must not disconnect from the grid but should achieve fault ride-through, i.e., maintain continuous operation. Disconnection from the grid is only permitted when the voltage amplitude exceeds the specified range. When the grid voltage experiences a sudden rise or fall due to a fault, the wind turbine should exchange reactive power with the grid to stabilize the grid voltage.

[0024] Specifically, when the terminal voltage of a direct-drive wind turbine drops due to a fault, the unit will enter a low-voltage ride-through state and return to normal operation after the fault ends. If the voltage drop is too large, the unit will directly disconnect from the grid. When the terminal voltage of a direct-drive wind turbine rises due to a fault, the unit will enter a high-voltage ride-through state. If this is not identified in real time during operation, it will trigger a cascading failure and cause grid collapse.

[0025] In summary, this invention discloses a method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network, comprising the following steps: Acquire electrical quantity data from offshore direct-drive wind farms, define the wind farm operating status and set status labels; associate the electrical quantity data and status labels to form a raw dataset and perform preprocessing; construct an improved one-dimensional convolutional neural network model, configure the model training parameters and perform dynamic training optimization; train the improved one-dimensional convolutional neural network model using the preprocessed raw dataset until the validation set accuracy converges; deploy the trained model to the wind farm monitoring system to process the electrical quantity data in real time and output the status identification results.

[0026] Specifically, the complexity of the simulation model is reduced by using feeder-based equivalent modeling, and the multi-scale feature extraction capability is enhanced by constructing an improved one-dimensional convolutional neural network model, thereby improving the accuracy of the training / test set. In particular, Gaussian noise, dilated convolution, and average pooling techniques are introduced into the improved one-dimensional convolutional neural network model, which effectively solves the gradient problem and insufficient multi-scale feature capture problem of traditional one-dimensional convolutional neural networks when dealing with complex wind farm status identification.

[0027] In this embodiment, acquiring electrical quantity data from a wind farm includes the following steps: A wind farm simulation model was built based on the PSASP platform. According to the wind farm topology, a preset number of wind turbines under the same feeder were equated to single virtual machine groups. Electrical quantity data of the feeder were collected by simulating grid faults and disturbances. In the equivalent settings of the wind farm simulation model, for N wind turbines under the same feeder, the electrical quantity X of a single unit was collected. single and feeder electrical quantity X line Verify the relation X line ≈N*X singleAfter its establishment, feeder data was adopted to replace stand-alone data.

[0028] Furthermore, the original dataset is preprocessed, including data cleaning, normalization, and noise-resistant data augmentation enabled only during the training phase; the data augmentation is used to simulate field noise, missing measurements, and electromagnetic interference to improve the model's robustness and generalization ability.

[0029] Furthermore, the noise-resistant data enhancement includes injecting colored noise (including white noise, pink noise, and Brownian noise) with a signal-to-noise ratio between 20 and 40 dB, power frequency and its harmonic interference, Gaussian noise, and impulse noise into different channels.

[0030] Furthermore, to maintain physical consistency across channels, the color noise, power frequency and its harmonic interference, Gaussian noise, and impulse noise are applied synchronously to all channels; the noise-resistant data enhancement is only effective on the training set after the training and validation sets are divided, and is turned off during the online inference phase.

[0031] Furthermore, based on voltage ride-through standards, the operating status of wind farms is defined, including normal state, high voltage ride-through state, low voltage ride-through state, low voltage ride-through recovery state, and grid disconnection state, such as... Figure 2 As shown; when the identification result is offline state, a tripping command is sent to the power grid control system; when the identification result is high voltage ride-through state or low voltage ride-through state, the reactive power compensation device control signal is activated.

[0032] In the specific implementation process, firstly, a wind farm simulation model is built in PSASP, including 33 direct-drive wind turbines and their feeders and main transformers, such as... Figure 5 As shown, G1-G33 are direct-drive wind turbines, connected to the infinite bus system S1 via three feeders and a main transformer. According to the defined wind farm operating conditions, each turbine has five operating states, resulting in 165 different states for the 33 turbines. Such a large number of classification labels hinders the algorithm from outputting accurate identification results. Considering that the parameters and operating states of direct-drive wind turbines on the same feeder are basically the same, the 11 turbines on the same feeder are treated as one turbine. This significantly reduces the number of classification labels in the algorithm and improves the accuracy of identification. The simulation system diagram after the equivalent treatment is shown below. Figure 7 As shown.

[0033] Secondly, the reliability of the equivalent model is verified by setting... Figure 5 A short-circuit fault occurred at the grid connection terminal of the wind farm shown. Electrical quantities at the generator terminals of feeder I and turbine Gl were collected. The electrical quantities at the generator terminals of Gl were magnified 11 times and compared with those of feeder I. The results are as follows. Figure 6As shown in the figure, the trends of electrical quantity changes are almost identical for both. Therefore, treating the 11 generating units as a single unit and using feeder electrical quantity data instead of generator terminal electrical quantity data for site identification can significantly reduce the number of classification labels and accurately characterize the operating status of the generating units below the feeder.

[0034] Considering that direct-drive wind turbines are generally in normal operation and the amount of actual fault test data available is small, we further use PSASP simulation to obtain terminal electrical quantity data under different operating conditions, including normal state, high voltage ride-through, low voltage ride-through, low voltage recovery and grid disconnection state.

[0035] Specifically, a direct-drive fan simulation system is built in PSASP as follows: Figure 8 As shown, G1 is a direct-drive wind turbine. According to the Wind Power Grid Connection Technical Regulations (Q / GDW 392-2009), the corresponding high / low voltage ride-through start and exit thresholds are set in PSASP. Different types of short-circuit faults are set on the high-voltage side of the step-up transformer adjacent to G1, with each fault lasting 0.1 seconds, to obtain electrical quantity data for the entire low-voltage ride-through process of the unit. Simultaneously, a sudden load shedding disturbance is set in the simulation system to obtain electrical quantity data for the entire high-voltage ride-through process of the unit. The unit's electrical quantity data before the fault or disturbance occurs are data under normal conditions. A severe fault at the turbine terminal can be set to obtain the turbine terminal electrical quantity data under off-grid conditions.

[0036] In this embodiment, the original dataset is preprocessed, including data cleaning, normalization, and noise enhancement. Data preprocessing transforms the raw data into a format suitable for model input, thereby reducing noise and redundant information in the data.

[0037] The original dataset is normalized and scaled to the range [0,1] to eliminate differences in dimensionality and numerical range between different features. This normalization is achieved by calling the `mapminmax` function, as shown below: ; To input the data into the convolutional neural network, the above data needs to be flattened into a four-dimensional tensor, with its dimensions set to [N]. dim ,1,1,M] and [N dim [,1,1,N], where N dim This is represented by the feature dimension of the data, where M and N represent the number of samples in the training and test data, respectively. This process is achieved by calling the reshape function.

[0038] In this embodiment, an improved one-dimensional convolutional neural network model is constructed, such as... Figure 4As shown, based on a one-dimensional convolutional neural network, Gaussian noise, dilated convolution, and average pooling techniques are introduced to optimize the traditional 1D-CNN, such as... Figure 3 The diagram shows the state recognition algorithm flowchart of the 3D convolutional neural network model. The overall architecture of the improved network still consists of an input layer, multiple convolutional modules, pooling layers, global average pooling layers, fully connected layers, and a classification layer.

[0039] First, Gaussian noise is used to enhance the model's robustness by connecting the input layer to a Gaussian noise layer, which is activated only during training. Injecting Gaussian noise into the input data to disrupt its smoothness serves as a form of "pre-training" regularization, forcing the model to learn more robust features and avoiding oversensitivity to minor changes in the input data, thus better adapting to noisy conditions in real-world applications. Gaussian noise is a type of noise whose probability density function follows a normal (Gaussian) distribution. Its probability density function can be expressed as: ; Where μ is the mean, σ is the standard deviation, used to control noise intensity, and σ 2 Let Variance be the variance.

[0040] Assuming the original data is x, the data after adding Gaussian noise can be represented as: .

[0041] Secondly, dilated convolution expands the receptive field and multi-scale feature extraction. Compared with traditional convolution operations, dilated convolution significantly expands the receptive field without adding additional parameters by introducing gaps controlled by the dilation rate into the convolution kernel.

[0042] The convolutional module includes at least one dilated convolutional layer, introducing a gap controlled by the dilation rate r within the convolutional kernel, satisfying: E eks =K+(K-1)*(r-1), E eks The effective receptive field size for dilated convolution; An adaptive dilation rate scheduling module is introduced, multiple dilated convolutional branches are set in parallel in the same convolutional layer, and a bandwidth prior constraint is introduced to dynamically adjust the dilation rate of the convolutional layer based on the bandwidth score of electrical quantity data. Furthermore, when the configuration model training parameters are dynamically trained and optimized, one of the above-mentioned main enhancements is randomly selected in each training batch and a small noise enhancement can be superimposed. The enhancement intensity parameter is adaptively sampled within a preset range to improve the recall rate for rare working conditions.

[0043] The operation of a dilated convolutional layer satisfies: The formula for calculating the output signal y is: , where r=2, the convolution kernel is w, the convolution kernel size K=3, the receptive field is expanded to 5, i is the output signal index, and j is the convolution kernel index; combined with the standard convolutional layer to form a multi-scale feature extraction chain.

[0044] Furthermore, an adaptive expansion rate scheduling module is introduced into the network, the operation of which satisfies: Multiple dilated convolutional branches are connected in parallel within the same convolutional layer to capture temporal features at different scales. The dilation rate of the branches is set to a value... The outputs of each branch are fused using channel-gated attention-weighted fusion to obtain the output of this layer; During training, the weights of the expansion rate are dynamically adjusted by optimizing the objective function. By calculating the bandwidth score for the electrical quantity data in each training mini-batch and integrating the bandwidth score into the calculation of channel gating attention, the weights of the dilated convolution branches are adaptively adjusted based on the bandwidth score, thereby enabling rapid response to changes in the wind turbine's operating status and achieving a rapid reaction from the stability control device.

[0045] Unlike typical one-dimensional convolutional neural network algorithms, dilated convolution is introduced in the second layer of the convolution module, which introduces a gap into the convolution kernel to expand the receptive field of the kernel. After convolution and pooling operations, GlobalAverage is used to perform global average pooling on the feature data to reduce the feature dimension and reduce the computational and storage requirements of the model.

[0046] After the convolutional module, average pooling is used to calculate the average value of local regions. This preserves the overall information of the data, smooths data features, and makes the model less sensitive to small fluctuations in the input data, thereby improving the model's efficiency and robustness. Average pooling calculates the average value of local regions as follows: ; Where W is the pooling window size and S is the step size.

[0047] Average pooling reduces the size of the feature map by calculating the average value of local regions. After the features are compressed by the global average pooling layer, they are connected to the fully connected layer. The output layer uses the Softmax activation function to generate a 5-dimensional state probability vector.

[0048] Furthermore, configuring model training parameters for dynamic training optimization includes the following steps: The Adam optimizer, known for its fast convergence and adaptive learning rate adjustment, is employed. The initial learning rate is 0.001, and a piecewise constant decay strategy is used, reducing the learning rate to 0.1 times its original value every 400 epochs. L2 regularization is used to constrain model parameters, with an L2 regularization coefficient of 1×10⁻⁴. This constraint on network weights prevents overfitting in the later stages of training. At the start of each training epoch, the training data is randomly shuffled, and the training process is visualized for real-time monitoring of progress and performance metrics. An early stopping mechanism is implemented: training stops when the accuracy fluctuation on the validation set is less than 0.1% for 10 consecutive epochs, based on real-time calculation of the validation set accuracy.

[0049] In convolutional neural networks (CNNs), higher accuracy can be achieved by increasing the number of convolutional layers and kernels. However, an excessive number of convolutional layers and kernels can lead to overfitting due to a significant increase in the model's training parameters, while also increasing training time and computational cost. Therefore, it is necessary to balance high accuracy and training time when constructing a CNN architecture.

[0050] In this embodiment of the invention, a three-layer convolutional module is set up. The first layer has 2×1 kernels, generating 16 convolutions, and the second and third layers have 2×1 kernels, generating 32 convolutions. To demonstrate the superiority of the improved one-dimensional convolutional neural network model, based on the above construction, Gaussian noise, dilated convolution, and average pooling are introduced into the one-dimensional convolutional neural network model to improve it. The results of a single iteration of the one-dimensional convolutional neural network model and the improved one-dimensional convolutional neural network model are shown below. Figure 9 and Figure 10 As shown in the figure, compared with the state recognition results of the one-dimensional convolutional neural network model, the improved one-dimensional convolutional neural network model has improved accuracy on both the training set and the test set.

[0051] To avoid the randomness of the results, the two state recognition programs were run 10 times, and the accuracy results of the one-dimensional convolutional neural network model and the improved one-dimensional convolutional neural network model were compared. Figure 11 and Figure 12 As shown, compared to the one-dimensional convolutional neural network model, the improved one-dimensional convolutional neural network significantly improves the accuracy of both the training and test sets, reaching no less than 98.5% and 97.5%, respectively. Specifically, the accuracy of the training set improves by an average of approximately 1.15%, and the accuracy of the test set improves by an average of approximately 0.75%. These results demonstrate that the improved one-dimensional convolutional neural network algorithm proposed in this invention is superior for identifying the operating status of wind farms.

[0052] Furthermore, in this embodiment of the invention, the input data may be subject to various interferences, such as environmental noise and signal distortion. If the model does not possess good robustness, it may make incorrect judgments under these circumstances. To ensure that the model can operate stably and reliably under various complex conditions, this embodiment of the invention verifies the robustness and generalization ability of the improved algorithm, examines its recognition rate on noisy signals, and adds 5%, 10%, 15%, and 20% Gaussian white noise to the divided test sets respectively, and performs experimental verification on each test set. Each test set is repeated 5 times, and the experimental results are shown in Table 1; the average accuracy of the 5 experimental test sets under different Gaussian noise levels is taken, and the comparison results are as follows: Figure 13 As shown.

[0053] Table 2. Diagnostic results of test samples under different Gaussian noise levels Depend on Figure 13 It can be seen that as the proportion of Gaussian noise increases, the diagnostic accuracy of the network model decreases. However, even with 20% white noise, it still maintains a high accuracy, indicating that the algorithm of this invention has good robustness and generalization ability.

[0054] In this embodiment, the real-time status identification of wind farm operation includes the following steps: The electrical quantity data collected in real time by the SCADA system of the wind farm is cleaned and normalized; the preprocessed electrical quantity data is input into the trained improved one-dimensional convolutional neural network model; and the state label corresponding to the maximum value of the Softmax output is taken as the recognition result.

[0055] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network, characterized in that, Includes the following steps: Acquire electrical quantity data of offshore direct-drive wind farms, define the wind farm's operating status, and set status labels; The electrical quantity data and the status label are associated to form a raw dataset, which is then preprocessed. An improved one-dimensional convolutional neural network model was constructed, and the model training parameters were configured for dynamic training optimization. The improved one-dimensional convolutional neural network model is trained using the preprocessed original dataset until the accuracy on the validation set converges. The trained model is deployed to the wind farm monitoring system to process electrical quantity data in real time and output status identification results.

2. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 1, characterized in that, The acquisition of electrical quantity data from wind farms includes the following steps: A wind farm simulation model was built based on the PSASP platform; Based on the wind farm topology, a preset number of wind turbines on the same feeder are equivalent to a single virtual machine group; By simulating power grid faults and disturbances, electrical quantity data of the equivalent feeder are collected.

3. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 2, characterized in that, In the equivalent settings of the wind farm simulation model, for N wind turbines on the same feeder, the electrical quantity X of each turbine is collected. single and feeder electrical quantity X line Verify the relation X line ≈N*X single After its establishment, feeder data was adopted to replace stand-alone data.

4. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 1, characterized in that, The original dataset is preprocessed, including data cleaning, normalization, and noise-resistant data augmentation enabled during the training phase. The noise-resistant data augmentation is used to simulate field noise, missing measurements, and electromagnetic interference. The noise reduction data enhancement includes injecting color noise, power frequency and its harmonic interference, Gaussian noise and impulse noise with a signal-to-noise ratio between 20 and 40 dB in different channels; The color noise, power frequency and its harmonic interference, Gaussian noise, and impulse noise are applied synchronously to all channels; the noise-resistant data enhancement is only effective on the training set after the training and validation sets are divided, and is turned off during the online inference phase.

5. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 1, characterized in that, The construction of the improved one-dimensional convolutional neural network model includes the following steps: The input layer is connected to a Gaussian noise layer, which is activated only during the training phase. The convolutional module includes at least one dilated convolutional layer, introducing a gap controlled by the dilation rate r within the convolutional kernel, satisfying: E eks =K+(K-1)*(r-1), E eks The effective receptive field size for dilated convolution; An adaptive dilation rate scheduling module is introduced, which sets multiple dilated convolutional branches in parallel on the same convolutional layer to capture temporal features at different scales. The dilation rate of the branches is set to a value of [value missing]. The outputs of each branch are fused by channel-gated attention weighting to obtain the output of this layer; During training, the weights of the inflation rate are dynamically adjusted by optimizing the objective function. Specifically, the bandwidth score is calculated for the electrical quantity data in each training mini-batch, and the bandwidth score is integrated into the calculation of the channel gating attention. The weights of the dilated convolution branches are adaptively adjusted based on the bandwidth score. After features are compressed by a global average pooling layer, they are fed into a fully connected layer. The output layer uses the Softmax activation function to generate a 5-dimensional state probability vector.

6. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 4, characterized in that, When the configuration model training parameters are dynamically trained and optimized, one of the above-mentioned main enhancements is randomly selected in each training batch and a small noise enhancement can be superimposed. The enhancement intensity parameter is adaptively sampled within a preset range to improve the recall rate for rare working conditions. The configuration model training parameters are dynamically trained and optimized, including the following steps: The Adam optimizer was used with an initial learning rate of 0.001, which decayed to 0.1 times the original value every 400 rounds. The model parameters are constrained by L2 regularization, and the L2 regularization coefficient is set to 1×10. -4 ; An early stop mechanism is set up, which calculates the accuracy of the validation set in real time and stops training when the accuracy fluctuation is less than 0.1% for 10 consecutive rounds.

7. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 5, characterized in that, The operation of the dilated convolutional layer satisfies: The formula for calculating the output signal y is: Where r=2, the convolution kernel is w, the convolution kernel size is K=3, the receptive field is expanded to 5, i is the output signal index, and j is the convolution kernel index; It is combined with standard convolutional layers to form a multi-scale feature extraction chain.

8. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 5, characterized in that, After the convolutional module, average pooling is used to calculate the average value of the local region. ； Where W is the pooling window size and S is the step size.

9. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 1, characterized in that, The real-time status identification of the wind farm's operating status includes the following steps: The electrical quantity data collected in real time by the SCADA system of the wind farm is cleaned and normalized. The preprocessed electrical quantity data is input into the trained improved one-dimensional convolutional neural network model. The state label corresponding to the maximum value of the Softmax output is taken as the recognition result.

10. The method for identifying the operating status of offshore direct-drive wind turbines based on an improved one-dimensional convolutional neural network according to claim 1, characterized in that, The operating status of a wind farm is defined based on the voltage ride-through standard, including normal state, high voltage ride-through state, low voltage ride-through state, low voltage ride-through recovery state, and grid disconnection state. When the identification result is that the grid is disconnected, a trip command is sent to the grid control system; When the identification result is a high voltage ride-through state or a low voltage ride-through state, the reactive power compensation device control signal is activated.

Images