Wind power time sequence anomaly detection method based on frequency domain interpolation and adaptive frequency attention
By introducing frequency domain interpolation and adaptive frequency attention mechanisms into wind power anomaly detection, combined with lightweight time-domain residual convolution, the problem of insufficient identification capability of non-stationary signals and periodic anomalies in wind farms is solved, realizing efficient online monitoring and intelligent operation and maintenance of wind farms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA SOUTHERN POWER GRID COMPANY
- Filing Date
- 2025-12-24
- Publication Date
- 2026-05-19
AI Technical Summary
Existing wind power anomaly detection methods have problems such as insufficient ability to identify non-stationary signals and periodic anomalies in wind farms, high model complexity and difficulty in real-time deployment, and sensitivity to noise interference.
By introducing frequency domain interpolation and adaptive frequency attention mechanism, the energy distribution characteristics of wind power time series data are explicitly modeled in the spectral space, and local feature refinement is performed by combining lightweight temporal residual convolution to generate anomaly scores.
It significantly improves the accuracy and robustness of wind power anomaly detection, is suitable for online monitoring and intelligent operation and maintenance of wind farms, reduces computational complexity and improves real-time performance.
Smart Images
Figure CN122065202A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent anomaly detection and time series analysis technology, specifically involving a wind power time series anomaly detection method based on frequency domain interpolation and adaptive frequency attention. Background Technology
[0002] With the rapid development of the global new energy industry, wind power, as an important component of green energy, has seen continuous growth in installed capacity and operational scale. Wind farms generate a large amount of time-series monitoring data during long-term operation, such as wind speed, wind direction, power output, voltage, and current. This data contains rich information about equipment operating status, but it is also affected by factors such as meteorological environment, equipment aging, and sensor drift, resulting in a large amount of non-stationarity, noise interference, and abnormal abrupt changes in the time-series signals. How to achieve high-precision, real-time anomaly detection of wind power time-series data in complex environments has become an important issue for the safe operation and intelligent maintenance of wind power. Traditional wind power anomaly detection methods mainly fall into two categories:
[0003] 1. Statistical and threshold-based methods, such as Z-score, ARIMA, and Hampel filtering. These methods rely on the assumption of data stationarity and have weak detection capabilities for nonlinear fluctuations and complex coupling relationships.
[0004] 2. Methods based on machine learning and deep learning, such as PCA, Isolation Forest, LSTM, AutoEncoder, and Transformer. These methods can learn the temporal patterns of data to some extent, but they often have the following problems:
[0005] (1) Insufficient utilization of frequency domain features makes it impossible to capture the energy distribution characteristics of periodic anomalies;
[0006] (2) The model is too complex or relies on large-scale parameters, making it difficult to deploy in real time at the edge nodes of wind farms;
[0007] (3) It is not sensitive enough to weak anomalies and multi-scale perturbations, and its detection performance is unstable.
[0008] In recent years, some studies have attempted to incorporate frequency domain information into time series analysis, such as using Fourier transform or wavelet decomposition to assist feature extraction. However, these methods typically only involve simple feature concatenation at the feature level, lacking a deep fusion mechanism. Furthermore, while traditional SRCNN networks possess good feature extraction capabilities, they are susceptible to noise interference and struggle to simultaneously model the collaborative relationship between the time and frequency domains in wind power anomaly detection scenarios. To address these issues, this invention proposes a wind power time series anomaly detection method based on frequency domain interpolation and adaptive frequency attention. This method introduces a frequency domain interpolation enhancement module on top of SRCNN, explicitly restoring the frequency distribution of anomaly features through fine-grained resampling in the spectral space. Simultaneously, an adaptive frequency attention (AFA) module is designed to dynamically adjust the weights of features in different frequency bands, thereby improving the model's response to multiple types of anomalies. Combining lightweight convolutional enhancement units and residual feature fusion strategies, this method can significantly improve the recognition performance of both periodic and non-periodic anomalies while maintaining low computational complexity.
[0009] A wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention includes the following steps:
[0010] Step 1 involves normalizing and frequency domain transforming the time series data, specifically including the following sub-steps:
[0011] Step 1-1, process the original time series data The input model aims to obtain the output sequence for the corresponding time interval through reconstruction or prediction. Used for calculating outlier scores.
[0012] Steps 1-2 involve calculating the mean and variance of the time series, and performing mean removal and variance normalization operations on the input series. To eliminate amplitude differences between different time series and improve model stability, layer normalization (LayerNorm) is then applied to the normalized sequence to further balance the feature distribution within the sequence.
[0013] Step 2 involves frequency domain feature extraction and attention weighting, which includes the following sub-steps:
[0014] Step 2-1: Normalize the sequence obtained in Step 1. Perform a Fast Fourier Transform (RFFT) to obtain the complex spectrum. The first k low-frequency real components are extracted and denoted as the dominant spectrum. To preserve key trend information and suppress high-frequency noise.
[0015] Step 2-2 introduces an adaptive frequency attention mechanism on the dominant spectrum, calculating weight coefficients through linear mapping and Sigmoid activation. And perform frequency domain weighting Here, ⊙ represents the frequency-wise product operation. This step enables dynamic importance allocation along the frequency dimension, allowing the model to focus on frequency regions strongly correlated with anomalies.
[0016] Step 3, perform trainable frequency domain interpolation and inverse transform reconstruction, which specifically includes the following sub-steps:
[0017] Step 3-1: Weight the spectrum using a trainable linear layer. Perform interpolation expansion to generate the target frequency length. The interpolation ratio is determined by the learnable parameter α = L y / L x The decision is made to implement an adaptive time length mapping.
[0018] Step 3-2: Pad the interpolated spectrum to obtain the complete output spectrum. And perform the inverse Fourier transform (IRFFT) to obtain the time-domain information. This serves as the initial result of the frequency domain reconstruction.
[0019] Step 4 involves refining local features using a temporal residual convolutional network, which includes the following sub-steps:
[0020] Step 4-1, take the output of step 3. A lightweight temporal residual module is input, and local contextual features are extracted using two layers of one-dimensional convolution (Conv1d-ReLU-Conv1d) to obtain the residual component R. t .
[0021] Step 4-2: Obtain the time-domain refinement result through additive fusion. It is used to correct transient losses or local anomaly smoothing problems caused by frequency domain interpolation, thereby improving the sensitivity to the identification of sudden anomalies.
[0022] Step 5 involves denormalizing the data and generating outlier scores, which includes the following sub-steps:
[0023] Step 5-1: Perform inverse normalization on the output sequence based on the mean and variance from Step 1 to restore it to the original numerical space.
[0024] Step 5-2: Calculate the mean squared error (MSE) between the model output and the actual observations, and take the error at the last time step as the anomaly score for the sample. A higher anomaly score indicates a greater reconstruction deviation at that moment, and a higher probability of an anomaly. Anomaly detection is ultimately achieved by setting thresholds or dynamic judgment rules.
[0025] As a further technical solution of the present invention, in step 2, the number of dominant frequencies k = L x / 2+1.
[0026] As a further technical solution of the present invention, the first layer of the temporal residual convolution module in step 4 has a kernel size of 3 and a channel number of 16, and the second layer has a kernel size of 3 and a channel number of 1, using the ReLU activation function to achieve nonlinear mapping.
[0027] Compared with the prior art, the present invention has the following advantages and effects:
[0028] This invention introduces frequency domain interpolation and an adaptive frequency attention mechanism to explicitly model the energy distribution characteristics of wind power time-series data in the spectral space, effectively solving the problem of insufficient ability of traditional methods to identify non-stationary signals, periodic anomalies, and weak anomalies. At the same time, it combines lightweight temporal residual convolution to refine and compensate local features, reducing information loss caused by noise interference and frequency domain reconstruction smoothing. Under the premise of ensuring low computational complexity, it significantly improves the accuracy, robustness, and real-time performance of anomaly detection, making it suitable for online monitoring and intelligent operation and maintenance scenarios of wind farms. Attached Figure Description
[0029] Figure 1 This is a flowchart of the method of the present invention.
[0030] Figure 2 This is a schematic diagram of the model of the present invention.
[0031] Figure 3 Visualization of the anomaly detection generation effect provided by the present invention (I).
[0032] Figure 4 Visualization diagram (II) of the anomaly detection generation effect provided by the present invention. Detailed Implementation
[0033] The technical solutions provided by the present invention will be described in detail below with reference to specific implementations. It should be understood that the following specific implementations are only used to illustrate the present invention and are not intended to limit the scope of the present invention.
[0034] This invention provides a wind power time-series anomaly detection method based on frequency domain interpolation and adaptive frequency attention. The method first performs mean-variance normalization on the wind power monitoring sequence and transforms it to the frequency domain, obtaining the main spectral features by extracting the main frequency band. Then, an adaptive frequency attention mechanism is introduced to weight the energy distribution of each frequency band, enabling the model to automatically focus on anomaly-sensitive frequency bands, thereby enhancing the response to anomaly frequency disturbances. Finally, a frequency domain interpolation module is used to perform spectral upsampling, and a time-domain signal is reconstructed and refined using inverse Fourier transform and time-domain residual convolution modules. An anomaly score is generated based on the reconstruction error, completing the wind power time-series anomaly detection.
[0035] Example: A wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention. The flowchart and model diagram of this invention are as follows. Figure 1 , Figure 2 As shown, it includes the following steps:
[0036] Step 1 involves normalizing and frequency domain transforming the time series data, specifically including the following sub-steps:
[0037] Step 1-1, process the original time series data The input model aims to obtain the output sequence for the corresponding time interval through reconstruction or prediction. Used for calculating outlier scores.
[0038] Steps 1-2 involve calculating the mean and variance of the time series, and performing mean removal and variance normalization operations on the input series. To eliminate amplitude differences between different time series and improve model stability, layer normalization (LayerNorm) is then applied to the normalized sequence to further balance the feature distribution within the sequence.
[0039] Step 2 involves frequency domain feature extraction and attention weighting, which includes the following sub-steps:
[0040] Step 2-1: Normalize the sequence obtained in Step 1. Perform a Fast Fourier Transform (RFFT) to obtain the complex spectrum. The first k low-frequency real components are extracted and denoted as the dominant spectrum. To preserve key trend information and suppress high-frequency noise.
[0041] Step 2-2 introduces an adaptive frequency attention mechanism on the dominant spectrum, calculating weight coefficients through linear mapping and Sigmoid activation. And perform frequency domain weighting Here, ⊙ represents the frequency-wise product operation. This step enables dynamic importance allocation along the frequency dimension, allowing the model to focus on frequency regions strongly correlated with anomalies.
[0042] Step 3, perform trainable frequency domain interpolation and inverse transform reconstruction, which specifically includes the following sub-steps:
[0043] Step 3-1: Weight the spectrum using a trainable linear layer. Perform interpolation expansion to generate the target frequency length. The interpolation ratio is determined by the learnable parameter α = L y / L x The decision is made to implement an adaptive time length mapping.
[0044] Step 3-2: Pad the interpolated spectrum to obtain the complete output spectrum. And perform the inverse Fourier transform (IRFFT) to obtain the time-domain information. This serves as the initial result of the frequency domain reconstruction.
[0045] Step 4 involves refining local features using a temporal residual convolutional network, which includes the following sub-steps:
[0046] Step 4-1: Take the output of step 3 A lightweight temporal residual module is input, and local contextual features are extracted using two layers of one-dimensional convolution (Conv1d-ReLU-Conv1d) to obtain the residual component R. t .
[0047] Step 4-2: Obtain the time-domain refinement result through additive fusion. It is used to correct transient losses or local anomaly smoothing problems caused by frequency domain interpolation, thereby improving the sensitivity to the identification of sudden anomalies.
[0048] Step 5 involves denormalizing the data and generating outlier scores, which includes the following sub-steps:
[0049] Step 5-1: Perform inverse normalization on the output sequence based on the mean and variance from Step 1 to restore it to the original numerical space.
[0050] Step 5-2: Calculate the mean squared error (MSE) between the model output and the actual observations, and take the error at the last time step as the anomaly score for the sample. A higher anomaly score indicates a greater reconstruction deviation at that moment, and a higher probability of an anomaly. Anomaly detection is ultimately achieved by setting thresholds or dynamic judgment rules.
[0051] In step 2, the number of dominant frequencies k = L x / 2+1.
[0052] In step 4, the temporal residual convolution module has a first-layer convolution kernel size of 3 and 16 channels, and a second-layer convolution kernel size of 3 and 1 channel. The ReLU activation function is used to implement nonlinear mapping.
[0053] Example:
[0054] Experimental conditions: All experiments were implemented using PyTorch on an NVIDIA RTX 3090 24GB GPU.
[0055] Experimental parameter settings: The model employs mean-variance-based normalization (RIN) and layer normalization at the input to improve training stability. The optimizer, Adam, is used to update model parameters. The initial learning rate is set to 0.001, and a learning rate decay strategy (StepLR, step size 5, decay coefficient 0.75) is used. During training, the L2 loss function (MSELoss) is used as the optimization objective, the batch size is set to 32, the maximum number of training epochs is 20, and an early stopping strategy (patience = 3) is used to prevent overfitting. In the testing phase, the squared reconstruction error at the last time step is used as the anomaly score to measure the degree of anomaly at each time point.
[0056] This paper evaluates the proposed anomaly detection model on two publicly available datasets, AIOPS and NAB, and compares its performance with five representative mainstream anomaly detection methods from the past three years, including AE, AR, TimesNet, OFA, and SRCNN. Model performance is primarily evaluated using F1, Precision, and Recall metrics, with higher values indicating better detection performance. Table 1 shows the performance comparison results of different models on various datasets. The best results are highlighted in bold, and the second-best results are highlighted with an underline. It can be seen that the proposed anomaly detection model based on frequency domain attention and temporal residual enhancement outperforms existing baseline methods in all comprehensive metrics, especially showing significant improvements in PR and F1 scores. This indicates that the model has stronger robustness and generalization ability in capturing potential frequency domain anomaly patterns and suppressing noise false alarms.
[0057] Table 1 Comparison of Model Prediction Performance
[0058]
[0059] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.
Claims
1. A wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention, characterized in that, Includes the following steps: Step 1: Normalize and frequency domain transform the time series. Step 2: Perform frequency domain feature extraction and attention weighting. Step 3: Perform trainable frequency domain interpolation and inverse transform reconstruction. Step 4: Refine local features using a temporal residual convolutional network. Step 5: Perform inverse normalization on the data and generate anomaly scores. Finally, anomaly detection is achieved by setting thresholds or dynamic judgment rules.
2. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention as described in claim 1, characterized in that, Step 1 involves normalizing and frequency domain transforming the time series data, specifically including the following sub-steps: Step 1-1, process the original time series data The input model aims to obtain the output sequence for the corresponding time interval through reconstruction or prediction. Used for anomaly score calculation Steps 1-2 involve calculating the mean and variance of the time series, and performing mean removal and variance normalization operations on the input series. To eliminate amplitude differences between different time series and improve model stability, layer normalization (LayerNorm) is then applied to the normalized sequence to further balance the feature distribution within the sequence.
3. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention as described in claim 1, characterized in that, Step 2 involves frequency domain feature extraction and attention weighting, which includes the following sub-steps: Step 2-1: Normalize the sequence obtained in Step 1. Perform a Fast Fourier Transform (RFFT) to obtain the complex spectrum. The first k low-frequency real components are extracted and denoted as the dominant spectrum. To preserve key trend information and suppress high-frequency noise, Step 2-2 introduces an adaptive frequency attention mechanism on the dominant spectrum, calculating weight coefficients through linear mapping and Sigmoid activation. And perform frequency domain weighting Where ⊙ represents the frequency-wise product operation, this step realizes the dynamic importance allocation of the frequency dimension, enabling the model to focus on the frequency regions that are strongly associated with anomalies.
4. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention according to claim 3, characterized in that, Step 3, perform trainable frequency domain interpolation and inverse transform reconstruction, which specifically includes the following sub-steps: Step 3-1: Weight the spectrum using a trainable linear layer. Perform interpolation expansion to generate the target frequency length. The interpolation ratio is determined by the learnable parameter α = L y / L x The decision is made to implement an adaptive time length mapping. Step 3-2: Pad the interpolated spectrum to obtain the complete output spectrum. And perform the inverse Fourier transform (IRFFT) to obtain the time-domain information. This serves as the initial result of the frequency domain reconstruction.
5. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention according to claim 4, characterized in that, Step 4 involves refining local features using a temporal residual convolutional network, which includes the following sub-steps: Step 4-1, take the output of step 3. A lightweight temporal residual module is input, and local contextual features are extracted using two layers of one-dimensional convolution (Conv1d-ReLU-Conv1d) to obtain the residual component R. t , Step 4-2: Obtain the time-domain refinement result through additive fusion. It is used to correct transient losses or local anomaly smoothing problems caused by frequency domain interpolation, thereby improving the sensitivity to the identification of sudden anomalies.
6. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention according to claim 5, characterized in that, Step 5 involves denormalizing the data and generating outlier scores, which includes the following sub-steps: Step 5-1: Perform inverse normalization on the output sequence based on the mean and variance from Step 1 to restore it to the original numerical space. Step 5-2: Calculate the mean squared error (MSE) between the model output and the actual observations, and take the error at the last time step as the anomaly score for the sample. The higher the anomaly score, the greater the reconstruction deviation at that moment, and the higher the probability of anomalies.
7. The time series anomaly detection method based on frequency domain interpolation and adaptive frequency attention according to claim 6, characterized in that, In step 2, the number of dominant frequencies k is equal to half the length of the input sequence plus one, i.e., k = L x / 2+1.
8. The wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention according to claim 7, characterized in that, The temporal residual convolution module in step 4 consists of two one-dimensional convolutional layers. The first one-dimensional convolutional layer has a kernel size of 3 and 16 output channels, used to extract local temporal features of the time series signal. The second one-dimensional convolutional layer also has a kernel size of 3 and 1 output channel, used to map the extracted features and generate a temporal residual signal. A ReLU activation function is introduced between the two convolutional layers to achieve nonlinear feature mapping, thereby enhancing the model's ability to express and model wind power time series anomaly patterns.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention as described in any one of claims 1 to 8.
10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instruction is executed by the processor, it implements the wind power timing anomaly detection method based on frequency domain interpolation and adaptive frequency attention as described in any one of claims 1-8.