Dimension reduction and equalization method for high-dimensional unbalanced data of wind turbine

Abstract

The invention discloses a dimension reduction and equalization method for high-dimensional unbalanced data of a wind turbine, and the method comprises the steps: removing singular values in a to-be-processed high-dimensional unbalanced data set, and obtaining an updated data set; performing normalization processing on the updated data set to obtain a normalized data set; performing oversampling on the icing data set in the normalized data set to obtain an oversampled data set; carrying out undersampling on an unfrozen data set in the normalized data set to obtain an undersampled data set; and merging the oversampling data set and the undersampling data set to obtain a merged data set, and performing standardization and dimension reduction processing on the merged data set in sequence to obtain a dimension-reduced balanced data set. The dimension reduction equilibrium data set obtained by the invention can effectively identify the early characteristics of icing of the wind turbine blade of the wind turbine generator, equipment maintenance can be carried out in time, the occurrence risk of disastrous accidents such as blade breakage and gearbox failure is reduced, stable operation and economic benefits of a wind power plant are ensured, and meanwhile, the difficulty of data processing and analysis is reduced.

Description

Technical Field

[0001] This invention relates to the field of wind turbine data processing technology, and in particular to a method for dimensionality reduction and balancing of high-dimensional unbalanced data of wind turbines. Background Technology

[0002] Against the backdrop of a global push for a clean energy transition, wind energy, as a clean and sustainable energy source, is playing an increasingly important role in the renewable energy sector. Wind power generation, with its unique advantages, has become a crucial component of energy structure adjustment. However, the practical application of wind power faces numerous challenges that urgently need to be addressed. Among these, the problem of wind turbine blade icing is particularly prominent. In some regions with severe winter weather conditions, wind turbine blades are highly susceptible to icing. This phenomenon not only significantly reduces power generation efficiency and drastically lowers the unit's output power but also significantly increases safety risks during operation, threatening equipment and personnel safety. Simultaneously, the frequent need for maintenance and equipment replacement significantly increases operating and maintenance costs. Therefore, accurately predicting wind turbine blade icing is of paramount importance for ensuring the safe and stable operation of wind power systems, improving wind energy utilization efficiency, and reducing operating costs.

[0003] Furthermore, Supervisory Control and Data Acquisition (SCADA) systems play a crucial role in the operation and management of large-scale wind farms. These systems provide vital support for ensuring the safe and efficient operation of wind turbines by collecting, transmitting, and processing their operational data in real time. However, SCADA systems currently face two challenging problems. Firstly, the collected data exhibits significant imbalance. Classifier models built upon this high-dimensional imbalanced data are prone to biased predictions, making it difficult to effectively identify early characteristics of wind turbine blade degradation. This can delay equipment maintenance, increase the risk of catastrophic accidents such as blade breakage and gearbox failure, and seriously threaten the stable operation and economic benefits of the wind farm. Secondly, with the continuous expansion of wind farm scale and the increasing refinement of monitoring needs, the dimensionality of data collected by SCADA systems has increased dramatically, exhibiting high-dimensional characteristics. High-dimensional imbalanced data not only increases the difficulty of data processing and analysis but also easily leads to problems such as excessively long model training times and overfitting, further reducing the predictive accuracy and reliability of the model. Summary of the Invention

[0004] Based on this, it is necessary to propose a method for dimensionality reduction and balancing of high-dimensional unbalanced data of wind turbines to address the above problems.

[0005] A method for dimensionality reduction and balancing of high-dimensional imbalanced data of wind turbines, wherein the high-dimensional imbalanced data is the operating data of the wind turbine generator, the method comprising: A high-dimensional imbalanced dataset to be processed is obtained, and singular values ​​in the dataset are removed to obtain an updated dataset. The high-dimensional imbalanced dataset to be processed consists of samples at multiple time points, and each sample contains multiple variables. The high-dimensional imbalanced dataset = [wind speed, wind direction angle, 25-second average wind direction angle, ambient temperature, x-axis acceleration, y-axis acceleration, active power, first pitch speed, second pitch speed, third pitch speed, first pitch power supply temperature, second pitch power supply temperature, third pitch power supply temperature, yaw position, yaw speed, generator speed, nacelle temperature, first pitch angle, second pitch angle, third pitch angle, first pitch motor temperature, second pitch motor temperature, third pitch motor temperature, first pitch power supply DC current, second pitch power supply DC current, third pitch power supply DC current]; The updated dataset is normalized to obtain a normalized dataset; The icing dataset in the normalized dataset is oversampled to obtain an oversampled dataset; the icing dataset consists of icing samples at multiple time points. The unfrozen dataset in the normalized dataset is undersampled to obtain an undersampled dataset; the unfrozen dataset consists of unfrozen samples at multiple time points. The oversampled dataset and the undersampled dataset are merged to obtain a merged dataset, and the merged dataset is standardized to obtain a standardized dataset. The standardized dataset is subjected to dimensionality reduction processing to obtain a dimensionality-reduced balanced dataset.

[0006] In one embodiment, removing singular values ​​from the high-dimensional imbalanced dataset to obtain an updated dataset includes: The same variable at different times in the high-dimensional imbalanced dataset to be processed is sorted in descending order to form a descending-order dataset; Determine the lower quartile and upper quartile of the descending dataset; The interquartile difference is determined based on the upper quartile and the lower quartile; The lower limit value is determined based on the interquartile difference and the lower quartile; The upper limit value is determined based on the interquartile difference and the upper quartile; Variables that are greater than the upper limit or less than the lower limit are considered singular values ​​and removed. The above operation is performed on each variable in the high-dimensional imbalanced dataset to be processed, and the variables that are not removed constitute the updated dataset.

[0007] In one embodiment, oversampling the icing dataset in the normalized dataset to obtain an oversampled dataset includes: A first number of real samples are randomly selected from the icing dataset; The generator generates a first number of fake samples based on random noise; The first output probability of the real sample and the second output probability of the fake sample are calculated by the discriminator. The discriminator loss is determined based on the first output probability and the second output probability; The discriminator network parameter set of the discriminator is backpropagated based on the discriminator loss to update the discriminator network parameter set; The generator loss is determined based on the second output probability; The generator network parameter set of the generator is backpropagated based on the generator loss to update the generator network parameter set in order to obtain the optimal generator. The pseudo-icing data is generated by the optimal generator, and the pseudo-icing data is added to the icing dataset to obtain the oversampled dataset.

[0008] In one embodiment, undersampling the unfrozen dataset in the normalized dataset to obtain an undersampled dataset includes: Configure a first memory and a second memory; Randomly select N% of the unfrozen data from the unfrozen dataset and store it in the first memory; the remaining unfrozen data is used as a pre-classification set. For N% of the unfrozen data in the first memory, the random forest algorithm is used for model training; The model is used to classify each sample in the unfrozen data. If the classification is correct, the sample is deleted; otherwise, the sample is stored in the second memory to form an undersampled dataset.

[0009] In one embodiment, the normalization process of the updated dataset to obtain a normalized dataset is achieved by the following expression: in, To update the dataset, This is a normalized dataset.

[0010] In one embodiment, the lower limit and upper limit values ​​for each variable are implemented using the following expression: in, This is the lower limit value; This is the upper limit value; It is the upper quartile; The lower quartile is denoted by IQR; the interquartile range is denoted by IQR. It is a constant.

[0011] In one embodiment, the discriminator loss and the updated discriminator network parameter set are solved by the following expression: in, For discriminator loss, This represents the first output probability of the real sample. This represents the second output probability of the fake sample. For the first quantity, For the discriminator network parameter set, Let be the gradient vector of the discriminator loss with respect to the set of discriminator network parameters.

[0012] In one embodiment, the generator loss and the updated generator network parameter set are solved by the following expression: in, For generator loss, This represents the second output probability of the fake sample. For the generator network parameter set, For the first quantity, The gradient vector of the discriminator loss with respect to the set of parameters of the generator network.

[0013] In one embodiment, the standardized data obtained by standardizing the merged dataset is achieved by the following expression: in, To standardize data, , ; For standardized datasets; For each sample in the merged dataset; This is the sample mean for each sample; For process variables The sampled values; Let be the standard deviation for each sample.

[0014] In one embodiment, the dimensionality reduction processing of the standardized dataset to obtain a dimensionality-reduced balanced dataset is achieved by the following expression: in, For standardized datasets; This is the correlation coefficient matrix; For the characteristic equation Solve for the eigenvalues. ; The contribution rate of variance Contribution rate to cumulative variance; For a dimensionality-reduced balanced dataset; coefficients For feature vectors The column vectors in To merge datasets Row vectors in the array.

[0015] The dimensionality-reduced balanced dataset obtained in this application can effectively identify early characteristics of icing on wind turbine blades, enabling timely equipment maintenance and reducing the risk of catastrophic accidents such as blade breakage and gearbox failure, thus ensuring the stable operation and economic benefits of the wind farm. It also reduces the difficulty of data processing and analysis. Attached Figure Description

[0016] 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] in: Figure 1 This is a diagram illustrating the application environment of a dimensionality reduction and equalization method for high-dimensional unbalanced data of wind turbines in one embodiment. Figure 2 This is a flowchart of a method for dimensionality reduction and equalization of high-dimensional imbalanced data of wind turbines in one embodiment; Figure 3 This is a schematic diagram of unfrozen data before singularity processing in one embodiment. Figure 4This is a schematic diagram of unfrozen data after singularity processing in one embodiment. Figure 5 This is a schematic diagram of icing data before singular value processing in one embodiment; Figure 6 This is a schematic diagram of icing data after singular value processing in one embodiment; Figure 7 This is a schematic diagram comparing experimental results in one embodiment; Figure 8 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Against the backdrop of a global push for a clean energy transition, wind energy, as a clean and sustainable energy source, is increasingly important in the renewable energy sector. Wind power generation, with its unique advantages, has become a crucial component of energy structure adjustment. However, the practical application of wind power faces numerous challenges that urgently need to be addressed. Among these, the problem of wind turbine blade icing is particularly prominent. In some regions with severe winter weather conditions, wind turbine blades are highly susceptible to icing. This phenomenon not only significantly reduces power generation efficiency and drastically lowers the unit's output power but also significantly increases safety risks during operation, threatening equipment and personnel safety. Simultaneously, the frequent need for maintenance and equipment replacement significantly increases operating and maintenance costs. Therefore, accurately predicting wind turbine blade icing is crucial for ensuring the safe and stable operation of wind power systems, improving wind energy utilization efficiency, and reducing operating costs. Furthermore, Supervisory Control and Data Acquisition (SCADA) systems play a key role in the operation and management of large-scale wind farms. These systems provide vital support for ensuring the safe and efficient operation of wind turbines by collecting, transmitting, and processing operational data in real time. However, SCADA systems currently face two challenging issues. On the one hand, the collected data exhibits significant imbalance. Classifier models built upon this imbalanced data are prone to biased predictions, making it difficult to effectively identify early characteristics of wind turbine blade degradation. This can delay equipment maintenance, increase the risk of catastrophic accidents such as blade breakage and gearbox failure, and seriously threaten the stable operation and economic benefits of wind farms. On the other hand, with the continuous expansion of wind farm scale and the increasing refinement of monitoring needs, the dimensionality of data collected by SCADA systems has increased dramatically, exhibiting high-dimensional characteristics. High-dimensional imbalanced data not only increases the difficulty of data processing and analysis but also easily leads to problems such as excessively long model training time and overfitting, further reducing the predictive accuracy and reliability of the model. To address the above technical problems, this application provides a method for dimensionality reduction and balancing of high-dimensional imbalanced data of wind turbines.

[0020] Figure 1 This diagram illustrates the application environment of a dimensionality reduction and equalization method for high-dimensional imbalanced data of wind turbines in one embodiment. (Refer to...) Figure 1This method for dimensionality reduction and equalization of high-dimensional imbalanced data from wind turbines is applied to a system for dimensionality reduction and equalization of such data. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. Terminal 110 is used to acquire a high-dimensional imbalanced dataset to be processed, and remove singular values ​​from the high-dimensional imbalanced dataset to obtain an updated dataset. The high-dimensional imbalanced dataset to be processed consists of samples at multiple time points, and each sample contains multiple variables. Server 120 is used to normalize the updated dataset to obtain a normalized dataset; oversample the icing dataset in the normalized dataset to obtain an oversampled dataset; the icing dataset consists of icing samples at multiple time points; undersample the non-icing dataset in the normalized dataset to obtain an undersampled dataset; the non-icing dataset consists of non-icing samples at multiple time points; merge the oversampled dataset and the undersampled dataset to obtain a merged dataset, and standardize the merged dataset to obtain a normalized dataset; perform dimensionality reduction processing on the normalized dataset to obtain a dimensionality-reduced balanced dataset.

[0021] like Figure 2 As shown, in one embodiment, a method for dimensionality reduction and equalization of high-dimensional imbalanced data of wind turbines is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The specific steps of this method for dimensionality reduction and equalization of high-dimensional imbalanced data of wind turbines are as follows: S10: Obtain the high-dimensional imbalanced dataset to be processed, and remove the singular values ​​in the high-dimensional imbalanced dataset to be processed to obtain the updated dataset. The high-dimensional imbalanced dataset to be processed consists of samples at multiple time points, and each sample contains multiple variables. The high-dimensional imbalanced dataset = [wind speed, wind direction angle, 25-second average wind direction angle, ambient temperature, x-direction acceleration, y-direction acceleration, active power, first pitch speed, second pitch speed, third pitch speed, first pitch power supply temperature, second pitch power supply temperature, third pitch power supply temperature, yaw position, yaw speed, generator speed, nacelle temperature, first pitch angle, second pitch angle, third pitch angle, first pitch motor temperature, second pitch motor temperature, third pitch motor temperature, first pitch power supply DC current, second pitch power supply DC current, third pitch power supply DC current]; S20: Normalize the updated dataset to obtain a normalized dataset; S30: Oversample the icing dataset in the normalized dataset to obtain an oversampled dataset; the icing dataset consists of icing samples at multiple time points. S40: Undersample the unfrozen dataset in the normalized dataset to obtain an undersampled dataset; the unfrozen dataset consists of unfrozen samples at multiple time points. S50: The oversampled dataset and the undersampled dataset are merged to obtain a merged dataset, and the merged dataset is standardized to obtain a standardized dataset; S60: Perform dimensionality reduction processing on the standardized dataset to obtain a dimensionality-reduced balanced dataset.

[0022] The dimensionality-reduced balanced dataset obtained in this application can effectively identify early characteristics of wind turbine blade degradation, thereby enabling timely equipment maintenance, reducing the risk of catastrophic accidents such as blade breakage and gearbox failure, and ensuring the stable operation and economic benefits of the wind farm. It also reduces the difficulty of data processing and analysis.

[0023] In one embodiment, the step S10 of removing singular values ​​from the high-dimensional imbalanced dataset to be processed to obtain an updated dataset includes: S101: Sort the same variable at different times in the high-dimensional imbalanced dataset to be processed in descending order to form a descending dataset; S102: Determine the lower quartile and upper quartile of the descending dataset; S103: Determine the interquartile difference based on the upper quartile and the lower quartile; S104: Determine the lower limit value based on the interquartile difference and the lower quartile; S105: Determine the upper limit value based on the interquartile difference and the upper quartile; S106: Treat variables that are greater than the upper limit value or less than the lower limit value as singular values ​​and remove them; S107: Perform the above operation on each variable in the high-dimensional imbalanced dataset to be processed, and the variables that are not removed constitute the updated dataset.

[0024] In one embodiment, the oversampling of the icing dataset in the normalized dataset in step S30 to obtain the oversampled dataset includes: S301: Randomly select a first number of real samples from the icing dataset; S302: The generator generates a first number of fake samples based on random noise; S303: Calculate the first output probability of the real sample and the second output probability of the fake sample using a discriminator; S304: Determine the discriminator loss based on the first output probability and the second output probability; S305: Backpropagate the discriminator network parameter set of the discriminator according to the discriminator loss to update the discriminator network parameter set; S306: Determine the generator loss based on the second output probability; S307: Backpropagate the generator network parameter set of the generator according to the generator loss to update the generator network parameter set in order to obtain the optimal generator; S308: Generate pseudo-icing data through the optimal generator, and add the pseudo-icing data to the icing dataset to obtain the oversampled dataset.

[0025] In one embodiment, the undersampling of the unfrozen dataset in the normalized dataset in step S40 to obtain the undersampled dataset includes: S401: Set the first memory and the second memory; S402: Randomly select N% of the unfrozen data from the unfrozen dataset and store it in the first memory; the remaining unfrozen data is used as a pre-classification set. S403: For N% of the unfrozen data in the first memory, the random forest algorithm is used for model training; S404: Use the model to classify each sample in the unfrozen data. If the classification is correct, delete the sample; otherwise, store the sample in the second memory to form an undersampled dataset.

[0026] In one embodiment, the normalization process performed on the updated dataset in step S20 to obtain a normalized dataset is achieved by the following expression: (1) in, To update the dataset, This is a normalized dataset.

[0027] In one embodiment, the Boxplot method is used to perform singular value detection on the 18-dimensional wind turbine data collected by the SCADA system, i.e., the high-dimensional imbalanced dataset to be processed. Boxplot is a statistical graph that displays the distribution of data, dividing the data statistics into five quantiles: lower limit value... Lower quartiles , median Upper quartiles With upper limit The Boxplot method can roughly show whether the high-dimensional imbalanced dataset to be processed is symmetric and whether the distribution is balanced. Sort the samples in the high-dimensional imbalanced dataset from largest to smallest and use the upper quartiles. The median is the value at the 75th percentile of the sample records. It is the value at the 50th percentile of the sample records, the lower quartile. The value at the 25th percentile of the sample records is used to calculate the difference between the upper and lower quartiles. This yields the interquartile range (IQR). A constant is then set. If the data is greater than or less than It is then determined to be a singular value. Generally, a value of 1.5 to 3 is used, representing a mild singularity. Take 1.5, extreme singular values The value is 3. The lower limit and upper limit for each variable in steps S104 and S105 are achieved using the following expressions: (2) (3) in, This is the lower limit value; This is the upper limit value; It is the upper quartile; The lower quartile is denoted by IQR; the interquartile range is denoted by IQR. It is a constant.

[0028] In one embodiment, from an icing dataset The first number of real samples are randomly selected from the middle. generator <strike> < / strike> A first number of fake samples are generated based on random noise. ; The discriminator loss in step S304 and the updated discriminator network parameter set in step S305 are solved by the following expressions: (4) (5) in, For discriminator loss, This represents the first output probability of the real sample. This represents the second output probability of the fake sample. For the first quantity, For the discriminator network parameter set, Let be the gradient vector of the discriminator loss with respect to the set of discriminator network parameters.

[0029] The generator loss in step S306 and the updated generator network parameter set in step S307 are solved by the following expressions: (6) (7) in, For generator loss, This represents the second output probability of the fake sample. For the generator network parameter set, For the first quantity, The gradient vector of the discriminator loss with respect to the set of parameters of the generator network.

[0030] In the above embodiment, a dynamic adversarial game is formed by iteratively executing the training steps of the discriminator and generator. In each training round: 1) Discriminator priority optimization: Update parameters based on real and fake samples to improve discrimination ability; 2) The generator is then optimized: parameters are updated based on the feedback from the discriminator to improve the ability of the generated samples to be indistinguishable from real ones.

[0031] This mechanism drives the two to evolve continuously in their competition, eventually approaching the optimal solution of the data distribution.

[0032] (2) Analysis of theoretical optimization objectives The core objective function of GAN is: in, Represents the expected value of the real sample. This represents the first output probability of the real sample. This represents the expected value of the fake sample. fake samples The second output probability. The generator attempts to minimize the objective function (making... The discriminator attempts to maximize the objective function (making...). and This forms a maximal and minima game.

[0033] In one embodiment, the standardized data obtained by standardizing the merged dataset in step S50 is achieved by the following expression: (8) (9) (10) (11) (12) in, To standardize data, , ; For standardized datasets; For each sample in the merged dataset; This is the sample mean for each sample; For process variables The sampled values; Let be the standard deviation for each sample.

[0034] In one embodiment, the dimensionality reduction processing of the standardized dataset to obtain a dimensionality-reduced balanced dataset in step S60 is achieved by the following expression: (13) (14) (15) in, For standardized datasets; This is the correlation coefficient matrix; For the characteristic equation Solve for the eigenvalues. ; The contribution rate of variance Contribution rate to cumulative variance; For a dimensionality-reduced balanced dataset; coefficients For feature vectors The column vectors in To merge datasets Row vectors in the array.

[0035] (1) The oversampling in this application first applies the SMOTE (Synthetic Minority Over-sampling Technique) algorithm, which generates new synthetic samples by analyzing the k nearest neighbors of minority class samples and performing linear interpolation, thereby alleviating the problem of data imbalance to a certain extent. However, the traditional SMOTE method is limited by the nature of its linear interpolation and is difficult to capture the complex distribution characteristics of real data. In order to overcome this limitation, this invention introduces the generator in the GAN (Generative Adversarial Networks) framework, which uses an adversarial training mechanism to learn the distribution pattern of real data and generates more complex and realistic nonlinear samples based on this. In the process of continuous optimization, the generator can create a more diverse sample set than the SMOTE method, effectively making up for the lack of minority class samples. At the same time, the discriminator is responsible for evaluating the quality of the generated samples, ensuring that these newly generated samples conform to the statistical characteristics of the original dataset, thereby improving the overall performance of the model. The SMOTE-GAN oversampling method cleverly combines the advantages of SMOTE and GAN technologies, which enhances the diversity and quality of minority class samples while also taking into account the issue of computational efficiency. Specifically, this method employs a phased strategy: first, SMOTE is used to quickly generate preliminary synthetic samples, and then GANs are used to further refine these samples, making them as close as possible to the real data distribution. This hybrid strategy not only improves the performance in handling imbalanced datasets but also provides new ideas and technical means for solving practical problems, possessing significant theoretical value and broad application prospects.

[0036] (2) The undersampling method of this application, namely CRF undersampling. This method aims to achieve a balance between categories and improve the prediction accuracy of the subsequent classification model for the minority class by intelligently filtering and reducing the number of majority class samples. Unlike traditional random undersampling methods, CRF undersampling uses random forests to evaluate the importance of each majority class sample, thereby more accurately identifying and retaining those samples that are crucial to maintaining the data structure. While reducing the number of majority class samples to balance the differences between categories, it improves the prediction accuracy of the subsequent classification model for the minority class.

[0037] (3) This application can identify and retain the most discriminative features while removing redundant and irrelevant features. This method not only improves computational efficiency but also enhances the model's performance when processing high-dimensional data. Furthermore, to address the problem of imbalanced datasets, SGCP introduces a resampling technique to adjust the ratio of positive to negative samples, bringing it closer to a balanced state. This technique includes both oversampling of minority class samples and undersampling of majority class samples, thereby ensuring that the trained classification model can more accurately predict minority class samples and improve overall classification performance. The SGCP method can effectively solve the problems of high dimensionality and imbalance in data, and can effectively improve the predictive performance and generalization of classification models.

[0038] As an experiment, this embodiment uses SCADA system monitoring data from two normally operating wind turbines (numbered 15# and 21#) in a wind farm in China as experimental samples (i.e., the high-dimensional unbalanced dataset to be processed). The high-dimensional unbalanced dataset to be processed contains 26 continuous numerical variables, covering multiple dimensions such as turbine operating parameters (e.g., grid-side active power, generator speed), environmental parameters (e.g., wind speed, ambient temperature), and state parameters (e.g., pitch motor temperature, bearing temperature). Specific parameters are detailed in Table 1. Table 2 shows the dimensionality-reduced balanced dataset.

[0039] Table 1. High-dimensional imbalanced dataset to be processed Table 2 Dimensionality Reduction Equilibrium Dataset The data labels include two categories: icy and non-iced, and their distribution is shown in Table 3. Icy samples account for less than 7%, exhibiting typical data imbalance characteristics. In this embodiment, to verify the synergistic performance of the four classifier models and the SGCP method, four comparative experiments were designed: Experiment 1 evaluates the predictive performance of these models on the same wind turbine data source; Experiments 2 and 3 further verify the predictive performance of these models applied to different wind turbines; Experiment 4 aims to verify whether the predictive performance is improved when multi-source data is added to the training set. Details of the data samples for each experiment are shown in Table 4. To address the imbalanced data problem, three preprocessing schemes were compared: the unprocessed SCADA dataset (A1), the dataset after random undersampling (A2), and the dataset processed by SGCP (A3).

[0040] Table 3 Dataset Labels and Data Distribution Table 4 Experimental Setup The Boxplot method was used to detect outliers in the 18-dimensional wind turbine SCADA data, where the outlier k was set to 3.

[0041] Figures 3-6 To process the results of wind turbine icing and non-icing data using the Boxplot method, Figure 3 This represents the unfrozen data before singular value processing. Figure 4 This represents the unfrozen data after singular value processing. Figure 5 This represents the icing data before singular value processing. Figure 6 This represents the icing data after singular value processing.

[0042] The results of outlier detection and resampling are shown in Table 5. The resampling results show that the ratio of icing data to normal data is close to 1:1.

[0043] Table 5 Comparison of original dataset, outlier detection processing, and resampling results Principal component analysis was performed on the resampled data, and 18 variables were extracted through dimensionality reduction. Table 6 shows the eigenvalues ​​and cumulative contribution rates calculated from the covariance matrices of these variables across all four experiments. Principal components with a cumulative contribution rate exceeding 90% were selected and used as input to the subsequent classification model. These principal components contained almost 90% of the information from the original 18 features. Table 6 shows that the first 8 principal components reached this threshold in Experiment 1, 9 in Experiment 2, 10 in Experiment 3, and 7 in Experiment 4.

[0044] Table 6. Cumulative contribution rate of 18 variables in the four experiments This embodiment uses three datasets, A1, A2, and A3, to validate the method through four comparative experiments. The experimental results are as follows: Figure 7 As shown in Table 7, the average evaluation metrics from the four experiments are summarized.

[0045] Four experiments were conducted using three datasets, A1, A2, and A3, and the results are as follows: Figure 7As shown in Experiment 1, A3 outperformed A1 on all four classifiers, demonstrating individual improvements of 12.1% (accuracy), 16.07% (precision), 21.56% (recall), and 28.82% (F1-score), resulting in an overall average improvement of 19.64%. Compared to A2, A3 improved accuracy by 6.56%, precision by 7.45%, recall by 5.86%, and F1-score by 7.44%, with an average improvement of 6.83%.

[0046] Experiment 2 shows that A3 outperforms other methods in both recall and F1-score, indicating that the SGCP method has better prediction performance and generalization ability. Experiment 3 shows that A3 has better recall performance, further confirming the predictive advantage of SGCP.

[0047] In Experiment 4, A3 showed a 12.35% improvement in accuracy, an 18.21% improvement in precision, a 22.48% improvement in recall, and a 29.64% improvement in F1-score compared to A1, with an average improvement of 20.67% across all metrics. Compared to A2, it showed a 7.96% improvement in accuracy, a 9.39% improvement in precision, a 6.74% improvement in recall, and a 9.10% improvement in F1-score, resulting in an overall performance advantage of 8.3%.

[0048] The average values ​​of the four experimental evaluation metrics are shown in Table 7, with higher ranking values ​​indicating better method performance. As shown in Table 7, A3 achieved the highest average ranking across all metrics in Experiments 1 and 4. This indicates that the SGCP classifier model exhibits better predictive performance on both single-source and multi-source datasets. In Experiments 2 and 3, A3 only performed better on recall and the F1-score for specific classification models, but its overall ranking was significantly lower than in Experiments 1 and 4. Therefore, these results may reveal some limitations of the method when handling heterogeneous datasets.

[0049] Table 7. Ranking of different datasets in the four classification models in the four experiments. Note: The numbers in bold indicate the best value for each row, and the numbers in parentheses indicate its ranking.

[0050] This invention relates to a data processing method (SGCP) for high-dimensional imbalanced data based on SCADA data, belonging to the field of wind power safety monitoring technology. Addressing the problem of insufficient prediction accuracy caused by the high dimensionality and sample imbalance of existing SCADA data, this method innovatively proposes a comprehensive data processing framework that effectively improves data quality through multi-dimensional feature engineering. The core innovations are: 1) proposing a staged data augmentation method combining SMOTE and GAN, generating high-quality and diverse minority class samples through the synergistic effect of linear interpolation (SMOTE) and nonlinear adversarial generation (GAN); 2) proposing a novel data-based undersampling method, CRF undersampling, which accurately retains key majority class samples by assessing sample importance; 3) proposing an end-to-end solution (SGCP) integrating feature selection and hybrid resampling to simultaneously solve the problems of high dimensionality and imbalance. Four sets of comparative experiments validated that this method demonstrates significant advantages on the A3 dataset. Compared to traditional methods A1 and A2, the key performance indicators (KPIs) are improved by an average of 19.64% and 6.83%, respectively, with Experiment 4 showing an even greater overall performance improvement of 20.67% and 8.3%. Particularly in the recall and F1-score dimensions, this method shows particularly outstanding optimization effects for specific classification models. Future technological extensions include developing adaptive processing mechanisms for heterogeneous data and constructing dynamic optimization frameworks to address multi-source data scenarios. The implementation of this invention will effectively enhance the anti-icing early warning capabilities of wind power generation systems, providing crucial technical support for ensuring the safe operation of wind turbines.

[0051] Figure 8 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 8 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for dimensionality reduction and equalization of high-dimensional imbalanced data from wind turbines. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a method for dimensionality reduction and equalization of high-dimensional imbalanced data from wind turbines. Those skilled in the art will understand that… Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0052] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for dimensionality reduction and balancing of high-dimensional unbalanced data of a wind turbine, wherein the high-dimensional unbalanced data is the operating data of the wind turbine generator, characterized in that, The method includes: Obtain the high-dimensional imbalanced dataset to be processed, and remove the singular values ​​in the high-dimensional imbalanced dataset to be processed to obtain the updated dataset; the high-dimensional imbalanced dataset to be processed consists of samples at multiple time points, and each sample contains multiple variables; the high-dimensional imbalanced dataset = [wind speed, wind direction angle, 25-second average wind direction angle, ambient temperature, x-direction acceleration, y-direction acceleration, active power, first pitch speed, second pitch speed, third pitch speed, first pitch power supply temperature, second pitch power supply temperature, third pitch power supply temperature, yaw position, yaw speed, generator speed, nacelle temperature, first pitch angle, second pitch angle, third pitch angle, first pitch motor temperature, second pitch motor temperature, third pitch motor temperature, first pitch power supply DC current, second pitch power supply DC current, third pitch power supply DC current]; The updated dataset is normalized to obtain a normalized dataset; The icing dataset in the normalized dataset is oversampled to obtain an oversampled dataset; the icing dataset consists of icing samples at multiple time points. The unfrozen dataset in the normalized dataset is undersampled to obtain an undersampled dataset; the unfrozen dataset consists of unfrozen samples at multiple time points. The oversampled dataset and the undersampled dataset are merged to obtain a merged dataset, and the merged dataset is standardized to obtain a standardized dataset. The standardized dataset is subjected to dimensionality reduction processing to obtain a dimensionality-reduced balanced dataset.

2. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 1, characterized in that, The process of removing singular values ​​from the high-dimensional imbalanced dataset to obtain the updated dataset includes: The same variable at different times in the high-dimensional imbalanced dataset to be processed is sorted in descending order to form a descending-order dataset; Determine the lower quartile and upper quartile of the descending dataset; The interquartile difference is determined based on the upper quartile and the lower quartile; The lower limit value is determined based on the interquartile difference and the lower quartile; The upper limit value is determined based on the interquartile difference and the upper quartile; Variables that are greater than the upper limit or less than the lower limit are considered singular values ​​and removed. The above operation is performed on each variable in the high-dimensional imbalanced dataset to be processed, and the variables that are not removed constitute the updated dataset.

3. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 2, characterized in that, Oversampling is performed on the icing dataset in the normalized dataset to obtain the oversampled dataset, which includes: A first number of real samples are randomly selected from the icing dataset; The generator generates a first number of fake samples based on random noise; The first output probability of the real sample and the second output probability of the fake sample are calculated by the discriminator. The discriminator loss is determined based on the first output probability and the second output probability; The discriminator network parameter set of the discriminator is backpropagated based on the discriminator loss to update the discriminator network parameter set; The generator loss is determined based on the second output probability; The generator network parameter set of the generator is backpropagated based on the generator loss to update the generator network parameter set in order to obtain the optimal generator. The pseudo-icing data is generated by the optimal generator, and the pseudo-icing data is added to the icing dataset to obtain the oversampled dataset.

4. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 1, characterized in that, The undersampling of the unfrozen dataset in the normalized dataset to obtain the undersampled dataset includes: Configure a first memory and a second memory; Randomly select N% of the unfrozen data from the unfrozen dataset and store it in the first memory; the remaining unfrozen data is used as a pre-classification set. For N% of the unfrozen data in the first memory, the random forest algorithm is used for model training; The model is used to classify each sample in the unfrozen data. If the classification is correct, the sample is deleted; otherwise, the sample is stored in the second memory to form an undersampled dataset.

5. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 1, characterized in that, The normalization process for the updated dataset, resulting in a normalized dataset, is achieved through the following expression: in, To update the dataset, This is a normalized dataset.

6. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 2, characterized in that, The lower limit and upper limit for each variable are achieved through the following expression: in, This is the lower limit value; This is the upper limit value; It is the upper quartile; The lower quartile is denoted by IQR; the interquartile range is denoted by IQR. It is a constant.

7. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 3, characterized in that, The discriminator loss and the updated discriminator network parameter set are solved by the following expression: in, For discriminator loss, This represents the first output probability of the real sample. This represents the second output probability of the fake sample. For the first quantity, For the discriminator network parameter set, Let be the gradient vector of the discriminator loss with respect to the set of discriminator network parameters.

8. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 3, characterized in that, The generator loss and the updated set of generator network parameters are solved by the following expression: in, For generator loss, This represents the second output probability of the fake sample. For the generator network parameter set, For the first quantity, The gradient vector of the discriminator loss with respect to the set of parameters of the generator network.

9. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 5, characterized in that, The standardized data obtained by standardizing the merged dataset is achieved through the following expression: in, To standardize data, , ; For standardized datasets; For each sample in the merged dataset; This is the sample mean for each sample; For process variables The sampled values; Let be the standard deviation for each sample.

10. The method for dimensionality reduction and equalization of high-dimensional unbalanced data of wind turbines according to claim 9, characterized in that, The dimensionality reduction process performed on the standardized dataset to obtain a dimensionality-reduced balanced dataset is achieved through the following expression: in, For standardized datasets; This is the correlation coefficient matrix; For the characteristic equation Solve for the eigenvalues. ; The contribution rate of variance Contribution rate to cumulative variance; For a dimensionality-reduced balanced dataset; coefficients For feature vectors The column vectors in To merge datasets Row vectors in the array.

Images