Wind speed data anomaly detection and correction method and system based on neural network

Abstract

The invention provides a wind speed data anomaly detection and correction method and system based on a neural network, and belongs to the technical field of wind speed data prediction.The method comprises the steps that historical wind speed data of a target wind power plant and historical wind speed data of a reference wind power plant highly associated with the target wind power plant are obtained; performing probability estimation on the historical wind speed data of the target wind power plant by adopting a self-detection model based on kernel density estimation to obtain a first detection result; performing wind speed data prediction of the target wind power plant on the historical wind speed data of the reference wind power plant by adopting an external detection model based on a neural network to obtain a second detection result; wherein the external detection model is a hybrid model combining variational mode decomposition, a convolutional neural network and a PFformer; analyzing the first detection result and the second detection result according to an analysis rule to obtain an abnormal data set, and performing data deviation correction on abnormal data. According to the invention, the accuracy is improved through dual model detection.

Description

Technical Field

[0001] This invention belongs to the field of wind speed data prediction technology, and in particular relates to a method and system for detecting and correcting wind speed data anomalies based on neural networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Against the backdrop of global sustainable development, the world's energy demand continues to rise due to accelerated industrialization and urbanization. Wind power, as a typical representative of new energy sources, has played an increasingly crucial role in the global energy structure transformation in recent years, thanks to its significant advantages such as cleanliness, environmental friendliness, and renewability. Its grid-connected installed capacity has maintained a steady growth momentum. For the wind power industry, high-quality, high-precision wind speed monitoring data is of paramount importance. It can not only effectively improve the accuracy and reliability of wind power forecasting, providing strong support for power dispatch and grid stability, but also provide solid and reliable data for key aspects such as identifying abnormal conditions in wind turbine units and diagnosing equipment faults, thus contributing to the efficient operation and sustainable development of the wind power industry.

[0004] Despite the rapid development of the wind power industry, wind farms face numerous challenges in actual operation. Wind farm sites are mostly concentrated in mountainous, desert, and near-shore areas. The complex terrain and harsh weather conditions, such as strong winds, torrential rains, freezing temperatures, and sandstorms, not only severely test the stability and durability of wind speed monitoring equipment but also affect the stability of signal transmission. Furthermore, long-term high-intensity operation accelerates wear and tear on instrument components, making them prone to mechanical failures. Lightning strikes and electromagnetic interference can also cause sensor malfunctions or signal transmission interruptions, easily generating abnormal monitoring data. Once this abnormal data is mixed with valid data, it interferes with the accuracy of data analysis, severely weakening the effectiveness and usability of the data. Currently, existing technologies include methods for detecting abnormal data by comparing the deviation between estimated and actual values; prediction methods combining BP neural networks and particle filtering algorithms; establishing meteorological observation data correction models by interpolating and fitting various meteorological anomaly data based on BP neural networks and sliding windows; and constructing a quality control algorithm based on a physical information bidirectional long short-term memory network, using the isolated forest algorithm to detect abnormal data in smoothed error sequences, thereby identifying abnormal wind speeds in the original time series. These methods have made some progress in meteorological data quality control, but some shortcomings remain. On the one hand, traditional anomaly detection methods generally rely on static threshold methods, which have low sensitivity and are difficult to effectively identify sudden abnormal data, while lacking a good correction mechanism for identified outliers. On the other hand, many detection methods do not consider the correlation between adjacent stations, focusing only on single-site detection scenarios and lacking collaborative modeling of multi-site data. Therefore, the models cannot capture the spatial dependencies between adjacent stations, limiting further improvements in anomaly detection accuracy. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a method and system for anomaly detection and correction of wind speed data based on neural networks. The system includes a self-detection component and an external detection component. The self-detection component utilizes historical wind speed monitoring data from the target wind farm and employs a probabilistic estimation method to detect anomalies. The external detection component uses reference sites highly correlated with the target wind farm, constructs a VMD-CNN-PFformer estimation model based on the screening results, and uses a threshold division method to detect anomalies in the target wind farm data. By combining the results of self-detection and external detection, anomaly data is accurately identified, and the wind speed anomaly data of the target wind farm is corrected, improving the overall quality and representativeness of the data.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention discloses a method for detecting and correcting anomalies in wind speed data based on neural networks, comprising: Obtain historical wind speed data for the target wind farm and reference wind farms that are highly correlated with the target wind farm; A self-detection model based on kernel density estimation is used to perform probability estimation on the historical wind speed data of the target wind farm to obtain the first detection result; An external detection model based on a neural network is used to predict the wind speed data of the target wind farm from the historical wind speed data of the reference wind farm, and a second detection result is obtained; wherein, the external detection model is a hybrid model combining variational mode decomposition, convolutional neural network and PFformer. The first and second detection results are analyzed according to the parsing rules to obtain an abnormal data set, and the abnormal data are corrected for data deviation.

[0007] Secondly, this invention discloses a wind speed data anomaly detection and correction system based on a neural network, comprising: The data acquisition module is used to acquire historical wind speed data of the target wind farm and a reference wind farm that is highly correlated with the target wind farm; The self-detection module is used to perform probability estimation on the historical wind speed data of the target wind farm using a self-detection model based on kernel density estimation, and obtain the first detection result; An external detection module is used to predict the wind speed data of the target wind farm from the historical wind speed data of the reference wind farm using an external detection model based on a neural network, and to obtain a second detection result; wherein, the external detection model is a hybrid model combining variational mode decomposition, convolutional neural network and PFformer. The detection and correction module is used to analyze the first detection result and the second detection result according to the parsing rules, obtain the abnormal data set, and correct the data deviation of the abnormal data.

[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned method for detecting and correcting anomalies in wind speed data based on neural networks.

[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned method for detecting and correcting anomalies in wind speed data based on neural networks.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for detecting and correcting wind speed anomalies. By analyzing self-detection results and external detection results, it ensures that all possible abnormal data are marked, thereby achieving comprehensive anomaly identification of wind speed data and subsequent deviation correction.

[0011] This invention constructs an anomaly detection model for wind speed prediction data. For anomaly wind speed data, it proposes a kernel density estimation model trained on wind farm data and a VMD-CNN-PFformer estimation model trained on highly correlated reference wind farm data. The kernel density estimation model is based directly on the inherent fluctuation range of the data and can accurately identify outliers that deviate from their own long-term statistical patterns, reducing misjudgments caused by short-term disturbances in the external environment. The VMD-CNN-PFformer estimation model effectively integrates external correlation information, captures spatial dependence, and utilizes the "regional meteorological consistency" law to make up for the limitations of single-site data, and can identify anomalies caused by local equipment failure rather than meteorological factors.

[0012] This invention uses Pearson correlation coefficients between different sites to screen wind farms that are strongly correlated with the target variable. Utilizing spatial correlation to screen strongly correlated sites effectively compensates for the limitations of information from a single site: their wind speed data reflects "regional meteorological trends," facilitating the identification of "anomalies caused by local equipment failures" and reducing noise interference such as transmission errors at single sites to avoid false detections. Simultaneously, data from strongly correlated sites improves the prediction accuracy of the VMD-CNN-PFformer model, enhances the reliability of anomaly detection benchmarks, and avoids residual threshold failures due to weak correlation with reference data. Furthermore, the consistency between the target site and the reference site can distinguish between regional meteorological events and local equipment failures.

[0013] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1 This is a flowchart of the wind speed data anomaly detection and correction method based on neural networks described in Embodiment 1 of the present invention.

[0016] Figure 2 This is a schematic diagram of the Pfformer model described in Embodiment 1 of the present invention.

[0017] Figure 3 This is a distribution diagram of the Pearson correlation coefficient as described in Embodiment 1 of the present invention.

[0018] Figure 4 This is a comparison chart of wind farm training and evaluation indicators as described in Embodiment 1 of the present invention.

[0019] Figure 5 This is a comparison chart of the results of the dual detection algorithm described in Embodiment 1 of the present invention and different algorithm models. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] Example 1 In one or more embodiments, a method for detecting and correcting anomalies in wind speed data based on neural networks is disclosed. A dual detection and correction method for wind speed data anomalies based on a hybrid model of kernel density estimation and VMD-CNN-PFformer is proposed. The self-detection part utilizes historical wind speed monitoring data of the target wind farm and fits its probability distribution based on kernel density estimation (KDE), using a probabilistic estimation method to detect anomalies. The external detection part uses Pearson correlation coefficient calculation to screen reference sites highly correlated with the target wind farm. Based on the screening results, a VMD-CNN-PFformer estimation model is constructed, and the model parameters and performance are optimized through cross-validation. A threshold division method is used to detect anomalies in the target wind farm data. Finally, the wind speed anomaly data of the target wind farm is corrected by combining the self-detection and external detection results. Figure 1 As shown, it includes the following steps: Step S1: Obtain historical wind speed data for the target wind farm and a reference wind farm that is highly correlated with the target wind farm.

[0024] Step S1-1: Calculate the correlation of wind speed data from different sites using the Pearson Correlation Coefficient (PCC) method. Based on the calculation results, accurately identify and select wind farms that have a significant correlation with the target variable as reference wind farms.

[0025] In order to select suitable wind farms and use their measured wind speed data as input variables for model training, this embodiment uses the Pearson correlation coefficient to select reference wind farms that meet the conditions.

[0026] It should be understood that the Pearson correlation coefficient is a statistical indicator that measures the linear relationship between two variables, and the strength of the relationship can be assessed through the covariance matrix of the data. Assume the two variables are... and The correlation coefficient is calculated as follows: (1) In the formula, Representing variables The sample mean, Representing variables The sample mean, Indicates the number of samples. The Pearson correlation coefficient reflects the strength of the linear correlation between two variables; the larger the absolute value, the stronger the correlation, as shown in Table 1. In wind speed data analysis for wind farms, the Pearson correlation coefficient can determine the correlation between wind speed data from different wind farms, thereby selecting wind farms with high correlation and using their wind speed data as training data.

[0027] In this embodiment, the training of the external detection model (VMD-CNN-PFformer) uses historical wind speed data from selected reference wind farms that are strongly correlated with the target wind farm as the core input. Reference sites with correlation coefficients exceeding 0.5 are selected by calculating the Pearson correlation coefficient, and their wind speed data are used as external correlation features for model training to learn the correlation between the wind speed at the target site and the wind speed at the reference sites, thereby improving prediction accuracy.

[0028] In subsequent steps, the trained external detection model still requires input of historical data from a reference wind farm during the prediction phase. Based on the wind speed data from the reference site and the correlation patterns learned during training, the model predicts and estimates the wind speed at the target site, then detects anomalies by comparing the residuals of the predicted and actual values. Inputting historical data from the reference site is essential for maintaining the logical consistency of the model's predictions, ensuring that the prediction results conform to the regional meteorological correlation characteristics.

[0029] Table 1. Explanation of Pearson Correlation Coefficient

[0030] In the selection of reference sites for wind farms, the Pearson correlation coefficient is more suitable for the characteristics of wind speed data than other techniques such as distance thresholding and mutual information. It can accurately capture the core feature of wind speed data, which is mainly linear correlation, and can directly quantify the correlation strength. It avoids the problems of distance thresholding, which may mistakenly select geographically close but meteorologically irrelevant sites, and mutual information, which reduces efficiency due to computational complexity and the introduction of redundant information. Its calculation only relies on basic statistics such as mean and covariance, which is simple and efficient. It can quickly process massive high-frequency time series data in batches, which is suitable for the real-time requirements of wind farms. It is superior to spatial clustering algorithms or deep learning embedding methods that require a lot of computing resources. Moreover, the results have clear physical meaning, and the selection criteria can directly correspond to the consistency of meteorological trends, which is convenient for model optimization and anomaly tracing. It overcomes the defects of "black box" and difficulty in interpreting the selection logic of techniques such as feature importance ranking based on neural networks.

[0031] Step S1-2: Obtain historical wind speed data for the target wind farm and a reference wind farm that is highly correlated with the target wind farm.

[0032] Step S2: Use a self-detection model based on kernel density estimation to perform probability estimation on the historical wind speed data of the target wind farm to obtain the first detection result.

[0033] During the self-detection phase, based on the historical wind speed monitoring data of the target wind farm, the probability distribution is fitted using kernel density estimation (KDE) technology. By comparing the relationship between the actual wind speed data of the target wind farm and the interval, the abnormal data can be accurately detected.

[0034] It should be understood that kernel density estimation, as a nonparametric probability estimation method, is widely used in processing data with nonlinear and non-Gaussian distribution characteristics. This model uses kernel functions to smooth and estimate the probability density of data, accurately describing continuous density functions. Its core advantage lies in the fact that it does not require any prior assumptions about the data distribution; instead, it analyzes the distribution patterns based on the inherent characteristics of the data itself, effectively avoiding the limitations of relying on specific distribution assumptions.

[0035] Specifically, the process of building a self-detection model includes: To construct a wind speed sample dataset for the target wind farm, historical wind speed monitoring data is obtained. A kernel density estimation model is constructed, and the probability density function of wind speed is fitted to the wind speed sample dataset of the wind farm using kernel density estimation. By smoothing each data point using a kernel function and then summing the effects of all data points, a continuous probability density estimate is finally obtained, expressed as: (2) In the formula, It is the sample size. This is a bandwidth parameter used to control the width and smoothness of the kernel function. It is a kernel function; Based on the fitting results, a 95% confidence interval is defined as the evaluation standard. Then, abnormal data is identified through probability estimation. If the actual wind speed data of the target wind farm exceeds the confidence interval, the first detection result is abnormal; otherwise, it is normal.

[0036] Step S3: Use an external detection model based on a neural network to predict the wind speed data of the target wind farm from the historical wind speed data of the reference wind farm, and obtain the second detection result; wherein, the external detection model is a hybrid model combining variational mode decomposition, convolutional neural network and PFformer.

[0037] In the external detection phase, a VMD-CNN-PFformer estimation model is constructed based on the selected reference sites that are highly correlated with the target wind farm. Cross-validation is used to optimize the model parameters and performance. Finally, anomaly detection of the target wind farm data is achieved through a threshold division method.

[0038] Specifically, in the external detection stage, historical wind speed data from a reference wind farm is used as input to the external detection model. This model performs predictive analysis on the wind speed data of the target site. An anomaly threshold is set based on the prediction results. If the prediction result exceeds the threshold, the second detection result is considered abnormal; otherwise, it is considered normal, thus achieving accurate identification of abnormal data. The anomaly threshold is determined by 1.05 times the maximum error between the predicted estimate and the original value in the validation set.

[0039] The external detection model is a hybrid model composed of Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), and Position-Free Transformer (PFformer). The specific model construction process includes: Step S3-1: Construct a variational mode decomposition unit to decompose the wind speed data into multiple modal components with different frequency characteristics. Each modal component typically corresponds to the fluctuation characteristics of the signal within a specific frequency range. This decomposition method can effectively reveal the multi-frequency components hidden in the signal, as expressed below: (3) (4) In the formula, Historical wind speed time series data for reference stations, Number of components For the first The center frequency of each component For the decomposition of the first One modal component, For the Dirac function, Let be the signal gradient, and t be the time t. Through optimization... and This ensures that the sum of all modal components approximates the original wind speed data as closely as possible. Furthermore, each component has a single frequency characteristic.

[0040] Furthermore, by introducing a penalty factor and Lagrange multipliers to solve the above constrained variational problem, the problem is transformed into an unconstrained variational problem: (5) In the formula, As a penalty factor for moderate bandwidth limitation, The Lagrange multiplier is used to constrain the fitting error. Solving constrained variational problems requires simultaneously satisfying constraints such as "the superposition of modal components equals the original signal" and "the frequencies of each component are orthogonal," which are mathematically difficult to solve directly using numerical methods. By introducing the penalty factor and the Lagrange multiplier, the constraints are transformed into part of the objective function, forming an unconstrained variational problem. This transformation allows for efficient solution through iterative optimization, facilitating the processing of massive time-series data in real-world wind power scenarios and improving computational feasibility.

[0041] Wind speed data from strongly correlated stations are complex signals, influenced by meteorological systems, topography, and temporal dynamics. The time series contains multiple frequency components: high frequencies correspond to instantaneous gusts or equipment noise, low frequencies correspond to daily / monthly average trends, and mid-frequency frequencies correspond to weather system changes. These components superimpose to form a nonlinear and non-stationary complex signal. Variational mode decomposition (VMD) is essential to break down the mixed-frequency signal into modal components dominated by single frequencies, clarifying the physical meaning of each component and facilitating targeted feature extraction by CNNs. Without VMD, the mixed frequencies make it difficult for the model to distinguish key features, reducing feature extraction accuracy and consequently affecting the prediction accuracy of the VMD-CNN-PFformer model for wind speed at the target station, causing the residual threshold for anomaly detection to fail.

[0042] This embodiment uses variational mode decomposition to decompose a complex signal into multiple mode components with different frequency characteristics. Each mode component typically represents the fluctuation of the signal within a specific frequency range. Thus, variational mode decomposition can help reveal the different frequency components in the signal, facilitating subsequent analysis or processing.

[0043] The output of VMD consists of multiple modal components. Each component is a wind speed subsequence dominated by a single frequency, belonging to the decomposed temporal features. These decomposition results are not directly input into the output layer of the CNN, but are first processed by decomposing each modal component... As an independent input convolutional layer for temporal features.

[0044] Step S3-2: Construct a convolutional neural network using a deep learning model, which is adept at processing time-series data.

[0045] In this embodiment, the input to the CNN is multiple modal components output by the Variational Mode Decomposition (VMD) unit. These modal components are time-series subsequences dominated by a single frequency obtained by decomposing historical wind speed data from strongly correlated reference wind farms using VMD. Each component represents the fluctuation of the wind speed signal within a specific frequency range. The output of the VMD unit is the direct input to the CNN, and the two have an "input-output" relationship. VMD decomposes the complex original wind speed signal into interpretable modal components, providing the CNN with clearer feature input, enabling the CNN to selectively extract wind speed fluctuation features within different frequency ranges.

[0046] The core structure of a CNN is "feature extraction-dimensionality reduction-integration," comprising an input layer, multiple convolutional and pooling layers, and a fully connected layer. Its hierarchy and connections are as follows: The input layer receives the modal components after VMD decomposition; following the input layer are convolutional layers, which perform local convolution operations on the input modal components using temporal convolution kernels to capture local correlation features of wind speed over a short period; each convolutional layer is followed by a pooling layer, which downsamples the feature map output by the convolutional layer, reducing data dimensionality while retaining key features; after multiple stacks of "convolutional-pooling layers," the feature map is flattened and input to the fully connected layer, where a weight matrix maps the high-dimensional features into a low-dimensional, discriminative feature vector. The final output of the CNN is a high-level feature vector after multi-layer feature extraction. This vector integrates key features of wind speed data at different time scales and serves as the input to the subsequent PFformer model for wind speed prediction of the target wind farm.

[0047] In this embodiment, the convolutional neural network model can effectively extract key features from wind speed data, specifically: The input is the VMD decomposition result of the first... Wind speed time series data for each modal component The convolution kernel is The convolution result is: (6) In the formula, The convolution feature at time t represents the weighted fluctuation of wind speed within a local time window, and s is the size of the convolution kernel window, which reflects the local time window of wind speed.

[0048] The wind speed characteristics output by the convolutional layer are as follows: The pooling window size is The max pooling result is: (7) The maximum wind speed characteristic within each pooling window is retained to highlight key meteorological events.

[0049] The wind speed feature vector output by the pooling layer is ( If the feature dimension is W, the weight matrix is ​​W, and the bias term is b, then the output of the fully connected layer is: (8) In the formula, The integrated wind speed feature vector incorporates key features from different frequency modes and is used as input to the PFformer model for wind speed prediction of the target wind farm.

[0050] In order to effectively extract key features from wind speed data, a convolutional neural network model is used for data processing. This model can automatically learn the spatial hierarchical structure of the data.

[0051] Step S3-3: Construct a multivariate time series forecasting model, using the Pfformer model, such as... Figure 2 As shown, it includes an input layer, an enhanced feature basic embedding module, a multi-head attention module, a feedforward module, an autoencoder-based embedding module, and a linear layer. The specific structure is as follows: the input layer is connected to the enhanced feature basic embedding module and the autoencoder-based embedding module respectively; the enhanced feature basic embedding module is connected in sequence to the first multi-head attention module, the first feedforward module, and the second multi-head attention module; the autoencoder-based embedding module is connected to the second multi-head attention module and the second linear layer respectively; the second multi-head attention module is connected in sequence to the third multi-head attention module, the second feedforward module, the first linear layer, and the summation layer; the second linear layer is connected to the summation layer, and the summation layer performs element-wise addition and then connects to the output layer.

[0052] In this embodiment, the multi-head attention module includes a multi-head attention layer and an addition and normalization layer; the feedforward module includes a feedforward layer and an addition and normalization layer.

[0053] It should be understood that PFformer is a Transformer-based model.

[0054] In this embodiment, the Enhanced Feature Basic Embedding (EFE) module in the PFformer model can combine the target time series variable with its related auxiliary time series variables, thereby effectively capturing the correlation between different series variables. The input target station wind speed sequence... and reference station wind speed sequence The two are concatenated along the time dimension, and then feature mapping is performed through a fully connected layer to generate a fused high-dimensional feature: (9) In the formula, This represents the concatenation of wind speed sequences from the target and reference stations along the time dimension. It is a fully connected layer, and its output is a high-dimensional feature vector that incorporates spatial correlation information.

[0055] The autoencoder-based embedding module (AEE) extracts latent states from time-series data, capturing long-term dependencies for future wind speed prediction. It consists of two parts: an encoder and a decoder. The latent representation generation formula is as follows: (10) In the formula, The input is time-series data (such as the feature vector output by a CNN). The extracted latent states. The future time-step predictions are represented as: (11) In the formula, for The potential state at any given moment For the predicted future Wind speed value at time step.

[0056] The VMD-CNN-PFformer model is used to obtain estimated wind speed data for the target site, and further, anomalies are determined based on the estimated values. The VMD-CNN-PFformer model takes wind speed data from a strongly correlated reference wind farm as input. After collaborative processing by Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), and PFformer, it outputs a predicted estimate of the wind speed at the target site. Then, it calculates the residual between the actual wind speed at the target site and this predicted estimate, comparing it to a preset anomaly threshold (determined by 1.05 times the maximum error between the predicted estimate and the original value in the validation set). If the absolute value of the residual exceeds the threshold, the data point is considered an anomaly; otherwise, it is considered normal. This process, combined with the spatial correlation of the reference site, identifies anomalies by the "degree of deviation between the predicted and actual values," effectively distinguishing between anomalies caused by equipment failure, transmission errors, etc., and normal meteorological fluctuations, thus improving the accuracy of anomaly detection.

[0057] The PFformer model addresses the challenge of traditional models failing to effectively capture complex dependencies between variables when dealing with rare or extreme events. By enhancing the representation of data features, this model effectively captures complex fluctuation patterns in wind speed data, thus optimizing data fitting accuracy.

[0058] The detection model is constructed through two main approaches: firstly, a probabilistic estimation model trained on the target wind farm's own data is built; secondly, a VMD-CNN-PFformer estimation model trained on highly relevant reference wind farm data is constructed. This hybrid model, combining Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), and PFformer, predicts and estimates wind speed data at the target site. Based on the prediction results, anomaly thresholds are determined to detect abnormal data, effectively improving wind speed prediction accuracy. The collaborative construction of these two models provides a solid foundation for subsequent data prediction. Finally, the detection results from both models are logically analyzed and integrated, and a comprehensive evaluation is conducted from multiple dimensions to measure the model's effectiveness and reliability.

[0059] Wind speed data is influenced by meteorological conditions and topography, often exhibiting nonlinear and non-Gaussian distribution characteristics. Kernel density estimation models do not require pre-setting the data distribution type; instead, they smooth historical data using a kernel function, directly fitting the probability density curve from the data itself. This characteristic avoids anomaly detection bias caused by incorrect distribution assumptions, making it particularly suitable for dynamically changing and complex data like wind speed. The model relies solely on the target site's own historical data, identifying anomalies through a 95% confidence interval. This approach, based directly on the inherent fluctuation range of the data, can accurately identify outliers deviating from their long-term statistical patterns, reducing misjudgments caused by temporary disturbances in the external environment.

[0060] The VMD-CNN-PFformer estimation model demonstrates several advantages in wind speed anomaly detection: First, it uses the Pearson correlation coefficient to select reference wind farms strongly correlated with the target site, using their wind speed data as input features to effectively integrate external correlation information, capture spatial dependence, and leverage the "regional meteorological consistency" principle to overcome the limitations of single-site data, thus identifying anomalies caused by local equipment failure rather than meteorological factors. Second, it collaboratively analyzes the complex features of wind speed data through multiple modules. VMD decomposes the original wind speed signal into multiple modal components of different frequencies, separating features such as high-frequency fluctuations and low-frequency trends, avoiding information loss from single-frequency analysis. CNN extracts modal components through convolutional layers. The model extracts local spatiotemporal features from the state components and reduces dimensionality through pooling layers, preserving key information while reducing computational complexity. PFformer then fuses target and reference site data through "enhanced feature embedding" and extracts temporal latent states using "autoencoder embedding," making it particularly adept at capturing the dependencies of extreme events and long-term series, thus addressing the problem of low prediction accuracy for rare anomalies in traditional models. Finally, the model calculates the "residual between predicted and actual values" by predicting wind speed at the target site. When the residual exceeds a preset threshold of 1.05 times the maximum error based on the validation set, it is considered an anomaly. This approach is sensitive to "sudden anomalies that deviate from the normal trend," compensating for the shortcomings of traditional static thresholding methods in adapting to dynamic changes.

[0061] Step S4: Analyze the first and second detection results according to the parsing rules to obtain the abnormal data set, and correct the data deviation of the abnormal data.

[0062] In this embodiment, the specific parsing rules are as follows: For the two sets of abnormal detection results—the first detection result generated by self-detection and the second detection result generated by external detection—analysis is performed according to the rules shown in Table 2. If neither the first nor the second detection result is abnormal, the abnormality level is judged as high, indicating a serious anomaly in the wind speed data; the data should be checked and corrected immediately. If the first detection result is normal and the second detection result is abnormal, the abnormality level is judged as medium, indicating a high probability of anomalies in the wind speed data, possibly due to equipment failure or external environmental interference causing transmission errors; certain inspection measures can be taken. If the first detection result is abnormal and the second detection result is normal, the abnormality level is judged as low, indicating some abnormal fluctuations in the wind speed data, possibly due to short-term weather changes or noise during the data acquisition process. It is recommended to check the stability of the data acquisition system. If both the first and second detection results are normal, the abnormality level is judged as normal, indicating that the wind speed data is normal and meets expectations.

[0063] Table 2. Dual-model anomaly detection evaluation and final result analysis

[0064] This embodiment integrates two detection results using a logical "OR" operation, and then calculates the recall rate to capture as many potential abnormal samples as possible, resulting in a final set of anomaly labels. Simultaneously, a logical "AND" operation is used to filter out anomaly samples that meet both detection criteria, and the precision rate is calculated to ensure the accuracy of the labeling results. Based on the integrated detection results, comprehensive anomaly identification of wind speed data can be performed, laying the foundation for subsequent data deviation correction.

[0065] Combining the anomaly marker results from self-detection and external detection, externally predicted wind speed values ​​are prioritized for imputation to ensure data smoothness and consistency with regional meteorological trends. The RMSE and coefficient of determination R of the corrected wind speed data are then calculated. 2 The effectiveness of the corrections will be evaluated and assessed.

[0066] This embodiment, by accurately identifying abnormal data and employing scientific methods for data repair and optimization, can significantly improve the overall quality and representativeness of the data. This provides a solid data foundation for the accurate construction of downstream wind power prediction models and the reliable operation of equipment fault diagnosis systems, thereby promoting the high-quality development of the wind power industry towards refinement and intelligence.

[0067] To comprehensively evaluate the performance of different models, this invention sets out multiple evaluation metrics. Among them, the performance evaluation metric for the VMD-CNN-PFformer model includes Mean Absolute Error (MAE). Root Mean Square Error (RMSE) The kernel density estimation model uses the miss coverage rate. The evaluation is based on indicators.

[0068] The accuracy of anomaly detection is ultimately evaluated using recall, precision, and F1 score. The formulas for each metric are as follows: (12) (13) (14) (15) (16) (17) In the formula, The average value of the target variable. For the sample size, It's a kernel function. and Given significance levels The lower and upper limits of the confidence interval at that time, the present invention Set to 0.05. TP (True Positive) represents the number of samples correctly predicted as positive by the model. FP (False Positive) represents the number of samples incorrectly predicted as positive by the model. FN (False Negative) represents the number of actual positive samples incorrectly predicted as negative by the model.

[0069] The accuracy and reliability of the model proposed in this invention are analyzed through practical examples. The data used in the examples of this invention are all from 11 onshore wind farms in the Shandong Peninsula, China.

[0070] 1) Data processing and selection of reference wind farms First, the wind speed data for each wind farm was split into a training dataset, a validation dataset, and a test dataset. Then, 100 data points were randomly selected from the test dataset and manually injected with an error margin of 5. The Pearson correlation coefficient was calculated between the various stations, such as... Figure 3 As shown, the Pearson correlation coefficients between various stations are used. Wind speed data from stations with a Pearson correlation coefficient exceeding 0.5 with the target station are selected as input variables for the external detection model.

[0071] 2) Detection of abnormal data In the self-testing section, the input variable of the kernel density estimation model is the historical data of the target element itself, and the output is the 95% confidence interval of the wind speed data of the target site.

[0072] In the external detection part, the VMD-CNN-PFformer model selects wind speed data from strongly correlated wind farms as input variables, and the output is the estimated value of wind speed data at the target site.

[0073] To optimize model performance, cross-validation was used for model tuning, and evaluation metrics were calculated for each fold, such as... Figure 4 As shown. Taking site 1 as an example, an evaluation index is calculated for each fold in the K folds, and the fold with the best index is selected as the optimal fold number of the model. The optimal fold number of the external detection model is the 5th fold, and the optimal kernel bandwidth of the self-detection model is 0.30. Finally, anomaly detection is carried out based on the test dataset.

[0074] Table 3 shows the confidence intervals and coverage rates of wind speed data for each station obtained through self-testing. Based on the self-testing evaluation results, the model demonstrates high accuracy in estimating the distribution intervals of wind speed monitoring data for each station at a 95% confidence level.

[0075] Table 3 Self-testing and evaluation index data table

[0076] Table 4 shows the maximum error, anomaly threshold, mean absolute error, and root mean square error of the final model of the site calculated after anomaly detection.

[0077] Table 4 External Testing Related Indicators

[0078] Based on the results of self-testing and external testing, the final recall, precision, and F1 score of the site were obtained, as shown in Table 5.

[0079] Table 5. Anomaly Detection Evaluation Index Data

[0080] 3) Correction of abnormal wind speed data Taking the anomaly detection results of stations 1 and 10 as examples, the wind speed values ​​predicted externally are prioritized for filling in the gaps, combining self-detection and external detection results to ensure data smoothness and consistency with regional meteorological trends. The RMSE and coefficient of determination R of the corrected wind speed data are then calculated. 2 The effectiveness of the corrections will be evaluated and assessed.

[0081] Table 6. Anomaly Correction Index Data

[0082] As shown in Table 6, the corrected RMSE for site 1 is 1.01, R 2 The RSE is 0.82, representing improvements of 34% and 22% respectively compared to the previous values. The revised RMSE for site 10 is 1.11. 2 The value was 0.70, representing an improvement of 28% and 23% respectively compared to the previous value. This indicates that anomaly detection and correction of wind speed monitoring data at wind farms can improve the quality of wind speed data at the sites.

[0083] 4) Comparative Experiment To verify the accuracy and stability of the method proposed in this invention, traditional anomaly detection methods, PFformer model, KDE model, and the detection method of this invention that does not use cross-validation were used as controls. The wind speed data of Wind Farm No. 1 was used as a sample to conduct comparative experimental analysis.

[0084] like Figure 5The diagram illustrates the calculation results of recall, precision, and F1 score for traditional anomaly detection methods, the KDE model, the PFformer model, a model without cross-validation, and the method of this invention in the examples presented in this invention. The results show that the method of this invention outperforms the control model in all indicators, verifying that the proposed method has superior performance and significantly improves upon the control model. In anomaly detection of wind speed monitoring data in wind farms, it demonstrates greater advantages compared to traditional anomaly detection methods, single detection methods, and models without cross-validation.

[0085] Example 2 In one or more embodiments, a wind speed data anomaly detection and correction system based on a neural network is disclosed, specifically including: The data acquisition module is used to acquire historical wind speed data of the target wind farm and a reference wind farm that is highly correlated with the target wind farm; The self-detection module is used to perform probability estimation on the historical wind speed data of the target wind farm using a self-detection model based on kernel density estimation, and obtain the first detection result; An external detection module is used to predict the wind speed data of the target wind farm from the historical wind speed data of the reference wind farm using an external detection model based on a neural network, and to obtain a second detection result; wherein, the external detection model is a hybrid model combining variational mode decomposition, convolutional neural network and PFformer. The detection and correction module is used to analyze the first detection result and the second detection result according to the parsing rules, obtain the abnormal data set, and correct the data deviation of the abnormal data.

[0086] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-mentioned method for detecting and correcting anomalies in wind speed data based on neural networks.

[0087] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for detecting and correcting anomalies in wind speed data based on neural networks.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A neural network-based wind speed data anomaly detection and correction method, characterized in that, The method comprises the following steps: obtaining historical wind speed data of a target wind farm and a reference wind farm highly associated with the target wind farm; using a self-detection model based on kernel density estimation to perform probability estimation on the historical wind speed data of the target wind farm to obtain a first detection result; using an external detection model based on a neural network to perform wind speed data prediction of the target wind farm on the historical wind speed data of the reference wind farm to obtain a second detection result; wherein the external detection model is a hybrid model combining variational modal decomposition, convolutional neural network and PFformer; analyzing the first detection result and the second detection result according to an analysis rule to obtain an abnormal data set, and correcting the abnormal data for data bias.

2. The neural network-based wind speed data anomaly detection and correction method of claim 1, wherein, The correlation of wind speed data of different sites is calculated by using the Pearson correlation coefficient method, and based on the calculation result, the wind farms with significant correlation with the target variable are accurately identified and selected as the reference wind farm. 3.The neural network-based wind speed data anomaly detection and correction method of claim 1, wherein, The construction of the self-detection model based on kernel density estimation is as follows: the kernel density estimation is used to fit the probability density function of wind speed sample data set of the wind farm, the kernel function is used to smooth each data point, and then the influence of all data points is accumulated to finally obtain a continuous probability density estimation; based on the fitting result, a 95% confidence interval is determined as the judgment standard, and then the abnormal data is identified through the probability estimation means; when the actual wind speed data of the target wind farm exceeds the confidence interval, the first detection result is abnormal, otherwise it is normal.

4. The neural network-based wind speed data anomaly detection and correction method of claim 1, wherein, The external detection model is a hybrid model combining variational modal decomposition, convolutional neural network and PFformer. First, a variational modal decomposition unit is constructed to decompose the wind speed data into multiple modal components with different frequency characteristics, and the modal components with different frequency characteristics are input into a convolutional neural network to extract time sequence features. A multivariate time series prediction model constructed by a PFformer model is used to predict the estimated value of the wind speed data of the target site.

5. The neural network-based wind speed data anomaly detection and correction method of claim 4, wherein, The multivariate time series prediction model comprises an input layer, an enhanced feature basic embedding module, a multi-head attention module, a feedforward module, an embedding module based on an autoencoder, and a linear layer. The specific structure is as follows: the input layer is connected to the enhanced feature basic embedding module and the embedding module based on the autoencoder; the enhanced feature basic embedding module is connected to the first multi-head attention module, the first feedforward module, and the second multi-head attention module in sequence; the embedding module based on the autoencoder is connected to the second multi-head attention module and the second linear layer; the second multi-head attention module is connected to the third multi-head attention module, the second feedforward module, the first linear layer, and the addition layer in sequence; the second linear layer is connected to the addition layer, and the addition layer is connected to the output layer after element addition.

6. The neural network-based wind speed data anomaly detection and correction method of claim 1, wherein, The analysis rule is as follows: when neither the first detection result nor the second detection result is abnormal, the abnormal level is determined to be high; when the first detection result is normal and the second detection result is abnormal, the abnormal level is determined to be medium. When the first detection result is abnormal and the second detection result is normal, the abnormality level is determined as a low level; if both the first detection result and the second detection result are normal, the abnormality level is determined as normal.

7. The neural network-based wind speed data anomaly detection and correction method of claim 1, wherein, The data bias correction specifically includes: Combining the abnormal flag results of self-detection and external detection, the wind speed value predicted externally is preferentially used for filling to ensure smooth data and compliance with regional meteorological trends; and the RMSE and coefficient of determination R of the corrected wind speed data are calculated 2 The correction effect is evaluated and judged. 8.A neural network-based wind speed data anomaly detection and correction system, characterized in that, The method comprises the steps of: a data acquisition module configured to obtain historical wind speed data of a target wind farm and a reference wind farm highly associated with the target wind farm; a self-detection module configured to perform probability estimation on the historical wind speed data of the target wind farm by using a self-detection model based on kernel density estimation to obtain a first detection result; an external detection module configured to perform wind speed data prediction of the target wind farm on the historical wind speed data of the reference wind farm by using an external detection model based on a neural network to obtain a second detection result; wherein the external detection model is a hybrid model combining variational modal decomposition, convolutional neural network and PFformer; a detection and correction module configured to analyze the first detection result and the second detection result according to an analysis rule to obtain an abnormal data set, and to perform data bias correction on the abnormal data.

9. An electronic device, comprising: A computer readable storage medium storing computer instructions, wherein the computer instructions are executed by a processor to complete the neural network-based wind speed data anomaly detection and correction method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer readable storage medium storing computer instructions, wherein the computer instructions are executed by a processor to complete the neural network-based wind speed data anomaly detection and correction method of any one of claims 1-7.

Images