Wind turbine generator state monitoring method based on multi-working-condition identification and space-time diagram network
By combining K-means clustering and spatiotemporal graph neural network models, the problem of unexplored feature correlations and spatiotemporal relationships in wind turbine condition monitoring has been solved, enabling accurate fault identification and early abnormal signal capture under complex operating conditions, thus improving the accuracy and reliability of monitoring.
Patent Information
- Application Number
- CN202610149778.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-01
AI Technical Summary
Existing wind turbine condition monitoring methods fail to effectively exploit the nonlinear correlations between sensor features and lack spatiotemporal relationship analysis, making it difficult to accurately capture abnormal signals during complex operating condition transitions or minor early faults, resulting in frequent false alarms and missed alarms.
The K-means clustering algorithm was used to segment the wind turbine operation data, a spatiotemporal graph neural network model was constructed, and spatial and temporal features were extracted by combining GAT and LSTM modules. The active power residual was smoothed by EWMA and a monitoring threshold was set for status monitoring.
It significantly improves the accuracy and reliability of wind turbine condition monitoring, reduces false alarm and false alarm rates, and enhances the model's adaptability and fault identification capabilities.
Smart Images
Figure CN121952807A_ABST
Abstract
Description
A Wind Turbine Condition Monitoring Method Based on Multi-Condition Identification and Spatiotemporal Graph Network Technical Field
[0001] This invention relates to the field of wind power monitoring technology, and more specifically to a method for monitoring the condition of wind turbine generators based on multi-condition identification and spatiotemporal graph networks. Background Technology
[0002] Wind turbines are the direct devices that convert wind energy into electrical energy. With the rapid development of the wind power industry, more and more wind turbines are being deployed in harsh and remote onshore or near-shore areas. Under these operating conditions, the key components of wind turbines are subjected to alternating loads and extreme weather conditions for extended periods. Once a failure occurs, not only is the scheduling of maintenance resources complex and costly, but the uncertainty of the environment often leads to prolonged maintenance cycles. Prolonged downtime due to failures will result in significant economic losses for wind power companies. Therefore, establishing an efficient and intelligent operational status monitoring system to achieve timely identification and early warning of potential wind turbine failures has become a key technological aspect for ensuring the safe operation of wind farms and reducing costs and increasing efficiency.
[0003] In existing technologies, the main methods for monitoring the operational status of wind turbines are data-driven detection methods. These methods first utilize sensor data such as temperature, pressure, speed, and power recorded by the wind turbine's SCADA (Supervisory Control and Data Acquisition) system, and then use machine learning models to uncover the underlying operational patterns to complete the operational status monitoring. To address the impact of changes in natural environmental factors such as wind speed and direction on the operational status of wind turbines, existing technologies have introduced a method of dividing operating conditions. By dividing the wind turbine's operating process into different state intervals, the interference of environmental nonlinear fluctuations on the monitoring model can be reduced.
[0004] However, existing condition monitoring methods based on operating condition division still have significant limitations: on the one hand, when performing feature analysis after operating condition division, existing technologies often ignore the complex nonlinear correlations between different sensor features of SCADA under each operating condition, failing to deeply explore the decoupling relationships between features; on the other hand, existing monitoring models usually only focus on time series modeling, lacking the fusion analysis of the spatial topology relationship and temporal dynamic characteristics of the internal structure of the wind turbine. This makes it difficult for existing models to accurately capture abnormal offset signals when facing complex operating condition switching or minor early faults, easily leading to false alarms or missed alarms, and the reliability and generalization ability of the methods are poor. Summary of the Invention
[0005] The purpose of this invention is to provide a wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network. This method can fully analyze the spatial correlation between features and decouple feature relationships. When facing complex condition switching or minor early faults, it can accurately capture abnormal offset signals, which significantly improves the identification accuracy and reliability of wind turbine operation condition monitoring.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] In a first aspect, the present invention provides a method for wind turbine condition monitoring based on multi-condition identification and spatiotemporal graph networks, the method comprising:
[0008] S1. Obtain wind turbine operation data and divide it into training set and test set for preprocessing; Based on clustering algorithm, divide the preprocessed wind turbine operation data into operating conditions, and divide the data into multiple operating condition subsets with different operating condition characteristics.
[0009] S2, construct the graph structure corresponding to each subset of working conditions and input it into the preset spatiotemporal graph neural network model, and train the spatiotemporal graph neural network model;
[0010] S3, use the trained spatiotemporal graph neural network model to predict the wind turbine test set data, obtain the active power prediction value, calculate the active power residual based on the active power prediction value and the actual active power value, and perform smoothing processing.
[0011] S4. Based on the training set, a monitoring threshold range is set, the number of times the active power residual exceeds the monitoring threshold range is counted, and the result is analyzed and compared with the preset warning value to obtain the wind turbine operating status monitoring result.
[0012] As a preferred embodiment of the present invention, the wind turbine operating data includes data from multiple sensors; the method for preprocessing the wind turbine operating data includes normalizing the multiple sensor data.
[0013]
[0014] Where, x i,j Let xj be the sensor data of the j-th sensor at time i, xjmin be the minimum value of the j-th sensor across all statistical times, xjmax be the maximum value of the j-th sensor across all statistical times, and xi,jnorm be the normalized sensor data of the j-th sensor at time i.
[0015] As a preferred embodiment of the present invention, in step S1, the clustering algorithm is specifically the K-means clustering algorithm, which uses the elbow method to determine the number of working condition features in order to divide the data into multiple working condition subsets.
[0016] As a preferred embodiment of the present invention, the spatiotemporal graph neural network model includes a GAT module for extracting spatial dependencies and an LSTM module for extracting temporal dynamic characteristics.
[0017] As a preferred embodiment of the present invention, in step S2, the loss function used to train the spatiotemporal graph neural network model is... for:
[0018]
[0019] Where N represents the total number of data values, This represents the actual value of active power. This represents the predicted value of active power.
[0020] As a preferred embodiment of the present invention, in step S3, the active power residual is smoothed using the EWMA smoothing method:
[0021]
[0022] Among them, E t With E t-1 The EWMA values at the current time t and the previous time are x and t, respectively. t Let be the active power residual at time t, and λ be the smoothing coefficient.
[0023] As a preferred embodiment of the present invention, the data types corresponding to the operating condition characteristics include: grid current, spindle speed, yaw angle, 30s average wind direction, 30s average wind speed, 30s average power, generator speed, generator winding temperature, generator drive end bearing temperature, generator non-drive end bearing temperature, nacelle temperature, and ambient temperature.
[0024] In a second aspect, the present invention also provides an electronic device, including a processor and a memory;
[0025] The processor is connected to the memory;
[0026] Memory, used to store executable program code;
[0027] The processor reads the executable program code stored in the memory and runs the program corresponding to the executable program code to perform the steps of the wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network described above.
[0028] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network.
[0029] In summary, the present invention has the following beneficial effects:
[0030] This invention utilizes the K-means clustering algorithm to classify the operating conditions of wind turbines, effectively solving the problem in existing monitoring methods where the model is easily affected by changes in operating conditions due to the mixing of data from different operating conditions. It constructs corresponding graph structures based on the operating characteristics of wind turbines under each operating condition and adjusts the spatiotemporal graph neural network model for each condition, ensuring that information from important nodes is aggregated onto prediction nodes. This guarantees that the model can still accurately aggregate information from important nodes when facing complex operating condition switching, significantly improving the method's adaptability.
[0031] In this invention, the spatiotemporal graph neural network model is constructed based on graph attention network (GAT) and long short-term memory network (LSTM), which can collaboratively extract the spatial dependency and temporal dynamic characteristics of wind turbine operation data. In addition, by combining power residual analysis after EWMA smoothing and monitoring threshold determination, this method can capture abnormal offset signals caused by minor early faults, reducing the incidence of false alarms and missed alarms caused by signal noise or environmental fluctuations, and has high practicality. Attached Figure Description
[0032] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 is a flowchart of the method of the present invention;
[0034] Figure 2 is a diagram illustrating the execution steps of a method according to an embodiment of the present invention;
[0035] Figure 3 is a schematic diagram of the K-means working condition classification results;
[0036] Figure 4 shows the cross-correlation coefficients among the characteristics of wind turbine units under different operating conditions;
[0037] Figure 5 shows the topology of the wind turbine unit without different operating conditions;
[0038] Figure 6 shows the structure of the wind turbine unit without different operating conditions;
[0039] Figure 7 shows the structure of the wind turbine under different operating conditions;
[0040] Figure 8 shows the comparison between the model-predicted active power and the actual active power under different operating conditions;
[0041] Figure 9 shows the status monitoring results of the wind turbine under different operating conditions. Detailed Implementation
[0042] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.
[0043] This embodiment provides a wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network, as shown in Figures 1 and 2. The specific steps of the method are as follows:
[0044] S1. Obtain wind turbine operation data and divide it into training and test sets for preprocessing. Based on clustering algorithm, divide the preprocessed wind turbine operation data into operating conditions, and divide the data into multiple operating condition subsets with different operating condition characteristics.
[0045] This embodiment uses an online SCADA database of a doubly-fed asynchronous wind turbine in a southwestern region to acquire wind turbine operating data. It extracts data from various sensors, including inputs, internal turbine information, and outputs. Focusing on the wind turbine generator components, the selected features after screening are: grid current, main shaft speed, yaw angle, average wind direction (30s), average wind speed (30s), average power (30s), generator speed, generator winding temperature, generator drive-end bearing temperature, generator non-drive-end bearing temperature, nacelle temperature, and ambient temperature – a total of 14 features. The operating condition characteristics of the wind turbine operating data are shown in Table 1 below.
[0046] Table 1
[0047] Next, training and testing datasets are created, and both datasets are preprocessed. Data preprocessing includes data cleaning and normalization. Data to be cleaned includes missing values, zero-power data, data from power-limited operations, and outlier data. Data normalization is then performed using the following formula:
[0048]
[0049] Where, x i,j Let xj be the sensor data of the j-th sensor at time i, xjmin be the minimum value of the j-th sensor across all statistical times, xjmax be the maximum value of the j-th sensor across all statistical times, and xi,jnorm be the normalized sensor data of the j-th sensor at time i.
[0050] Then, the preprocessed training and test datasets are input into the K-means clustering algorithm model. The K-means clustering algorithm determines the number of operating conditions using the elbow method, as shown in Figure 3, and divides the running data into multiple subsets of operating conditions with similar operating characteristics.
[0051] The core evaluation metric for the elbow method is the sum of squared errors (SSE). This embodiment provides one method for its calculation:
[0052]
[0053] In the formula, K is the number of cluster centers (i.e., the number of divisions for the target working condition), and C i Let x be the number of samples in the i-th cluster. j Let m be the j-th sample point in the cluster. i Let |x| be the cluster center of the i-th cluster. j -m i || 2 Let be the Euclidean distance from each sample point in the i-th cluster to its cluster center.
[0054] S2, construct the graph structure corresponding to each subset of working conditions and input it into the preset spatiotemporal graph neural network model, and train the spatiotemporal graph neural network model.
[0055] After completing the operating condition segmentation and obtaining the operating condition subsets, it is necessary to construct a corresponding graph structure for each operating condition. In this embodiment, the graph structure is constructed as follows: First, the selected operating condition features are divided into two categories: one is temperature features, and the other is features that directly participate in the process of wind turbines converting wind energy into electrical energy. When the correlation coefficient between temperature features is greater than 0.6, they are connected; and when the cross-correlation coefficient between the two types of features is greater than 0.5, they are connected. In this embodiment, the cross-correlation coefficient between operating condition features is shown in Figure 4.
[0056] As shown in Figure 5, the wind turbine topology structure is divided into different colors to represent different information categories. The different information categories are connected based on real physical relationships. Then, based on the analysis of the turbine topology structure, the basic diagram structure of the wind turbine without division of operating conditions is obtained, as shown in Figure 6.
[0057] As shown in Figure 6, taking a wind turbine without defined operating conditions as an example: First, based on the operating mechanism of the wind turbine, the physical connection relationships between key components are predetermined, and the corresponding characteristic nodes are connected. The wind turbine receives wind energy through the rotor, which drives the main shaft to rotate. Simultaneously, to improve the efficiency of wind energy reception, a yaw system is needed to ensure the rotor always faces the wind direction. Therefore, nodes 1, 2, 3, and 4 in Table 1 are interconnected. Since the main shaft drives the generator, nodes 3 and 5 are related and influence each other. Finally, the generator outputs electrical energy; therefore, node 5 points to node 6 in the diagram. Then, point to node 11 to complete the energy conversion from wind energy to mechanical energy and then to electrical energy. Subsequently, calculate the cross-correlation coefficient between the features under each operating condition, and connect the edges of the remaining temperature nodes (nodes 7, 8, 9, 10, and 12) according to the magnitude of the correlation (as shown in Figure 3). Since the generator is located in the nacelle, in order to reduce the influence of the environment on the temperature features of the components, all temperature features inside the generator should be reduced by the nacelle temperature, while the ambient temperature only affects the nacelle. Therefore, only node 13 and node 14 are connected to obtain the graph structure of the wind turbine unit without different operating conditions, as shown in Figure 6.
[0058] The formula for calculating the cross-correlation coefficient is as follows:
[0059]
[0060]
[0061] In the formula, ρ xy (k) is the cross-correlation coefficient, Rk xy(k) is the cross-correlation function, x and y represent two sets of feature sequences, N represents the sequence length, x(t) is the x value at time t, y(t+k) is the y value at time t+k, and k represents the lag value of the function.
[0062] Repeat this step for each working condition to obtain the corresponding graphical structure for each working condition, as shown in Figure 7.
[0063] The graph structures constructed for each working condition in the training set are input into the corresponding spatiotemporal graph neural network model. In this embodiment, the spatiotemporal graph neural network model includes a GAT module for extracting spatial dependencies and an LSTM module for extracting temporal dynamic characteristics.
[0064] For different operating conditions, corresponding spatiotemporal graph neural network models are trained to learn the normal operating mode of wind turbines under each condition.
[0065] The specific training steps are as follows:
[0066] The input sequence length and time step parameter of the dataset are pre-defined, where the input sequence length is determined by a sliding window and the time step parameter is used to characterize the time interval between adjacent samples; based on the sliding window size and the time step parameter, the original data is divided to obtain multiple data subsets of the same size;
[0067] Configure the number of layers, hidden layer dimensions, number of multiple attention heads, number of hidden layer neurons, output layer dimensions, training epochs, learning rate, optimization algorithm, and loss function of the GAT module in the model;
[0068] In single-head attention, the attention coefficient e between different nodes in the GAT module. ij and the normalized coefficient a ij It can be represented as:
[0069]
[0070] In the formula, a T N is a vector parameter. i For nodes The number of adjacent nodes, W is the shared weight matrix, and the weights are e ij x represents the importance of node j to node i. i and x j This represents the values of node i and node j at the current time.
[0071]
[0072] After passing through a multi-head attention GAT layer, the node's output is:
[0073]
[0074] In the formula, akij is the normalized coefficient after the calculation of the k-th attention, and Wij k M represents the shared weight matrix participating in the k-th attention calculation, where M is the number of attention heads and σ represents the non-linear activation function.
[0075] Under each operating condition, the data is input into the model according to the constructed graph structure, and the predicted active power value is output, as shown in Figure 8.
[0076] The loss value is calculated based on the predicted active power value, and the model parameters are updated using the Adam optimization algorithm.
[0077] The loss function used in the spatiotemporal graph neural network model is... for:
[0078]
[0079] In the formula, N represents the total number of data values. Represents the true value. This represents the predicted value.
[0080] Based on the time step, select the next subset of data and perform the above training steps until the training cycle reaches the set number of times.
[0081] S3. Use the trained spatiotemporal graph neural network model to predict the test set data of the wind turbine to obtain the predicted active power value. Calculate the active power residual based on the predicted active power value and the actual active power value and perform smoothing processing.
[0082] The trained model is used to predict active power, and the residual is calculated based on the predicted value and the actual value, and then smoothed.
[0083] S4. Based on the training set, a monitoring threshold range is set, the number of times the active power residual exceeds the monitoring threshold range is counted, and the results are analyzed and compared with the preset warning value to obtain the wind turbine operating status monitoring results.
[0084] For different operating conditions, threshold lines for the health status are set separately. The 3σ criterion in statistics is used as the judgment standard to determine the upper and lower threshold lines, including the following steps:
[0085] First, the residuals are calculated and smoothed. The calculation formula is as follows:
[0086]
[0087] In the formula, Et With E t-1 The EWMA values at the current time t and the previous time are x and t, respectively. t Let be the residual at time t, and λ be the smoothing coefficient.
[0088] The test set residual data is transformed into a normal distribution using quantile transformation.
[0089] The upper and lower thresholds of the transformed data are calculated using the 3σ principle:
[0090]
[0091]
[0092] Where UCL and LCL are the upper and lower thresholds, respectively, μ is the mean of the residuals, σ is the standard deviation of the residuals, λ is the smoothing coefficient, K is the control coefficient, and n is the sample length.
[0093] Based on the upper and lower threshold lines, a monitoring threshold range is obtained. The test dataset is input into the model trained for each operating condition, the predicted value is output and the residual is calculated. The residual is also smoothed. When the residual exceeds the monitoring threshold range multiple times in a row, the wind turbine is determined to be in an abnormal operating state and an early warning is triggered. Figure 9 shows a specific state monitoring result corresponding to this embodiment.
[0094] Several embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for wind turbine condition monitoring based on multi-condition identification and spatiotemporal graph networks, characterized in that, The method includes: S1, acquiring wind turbine operating data and dividing it into training and test sets, and preprocessing it; using a clustering algorithm to divide the preprocessed wind turbine operating data into operating conditions, dividing the data into multiple operating condition subsets with different operating condition characteristics; S2, constructing a graph structure corresponding to each operating condition subset and inputting it into a preset spatiotemporal graph neural network model, and training the spatiotemporal graph neural network model; S3, using the trained spatiotemporal graph neural network model to predict the wind turbine test set data, obtaining the predicted active power value, calculating the active power residual based on the predicted active power value and the actual active power value, and performing smoothing processing; S4, setting a monitoring threshold range based on the training set, counting the number of times the active power residual exceeds the monitoring threshold range, and analyzing and comparing it with a preset warning value to obtain the wind turbine operating status monitoring result.
2. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network according to claim 1, characterized in that, The wind turbine operating data includes data from multiple sensors; the method for preprocessing the wind turbine operating data includes normalization processing of the multiple sensor data. , where x i,j Let xj be the sensor data of the j-th sensor at time i, xjmin be the minimum value of the j-th sensor across all statistical times, xjmax be the maximum value of the j-th sensor across all statistical times, and xi,jnorm be the normalized sensor data of the j-th sensor at time i.
3. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network according to claim 1, characterized in that, In step S1, the clustering algorithm is specifically the K-means clustering algorithm. The K-means clustering algorithm determines the number of working condition features by using the elbow method to divide the data into multiple working condition subsets.
4. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network according to claim 1, characterized in that, The spatiotemporal graph neural network model includes a GAT module for extracting spatial dependencies and an LSTM module for extracting temporal dynamic characteristics.
5. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network according to claim 4, characterized in that, In step S2, the loss function used to train the spatiotemporal graph neural network model for: Where N represents the total number of data values, This represents the actual value of active power. This represents the predicted value of active power.
6. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network according to claim 1, characterized in that, In step S3, the active power residual is smoothed using the EWMA smoothing method: , of which E t With E t-1 The EWMA values at the current time t and the previous time are x and t, respectively. t Let be the active power residual at time t, and λ be the smoothing coefficient.
7. The wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network as described in claim 1, characterized in that, The data types corresponding to the operating condition characteristics include: grid current, spindle speed, yaw angle, 30s average wind direction, 30s average wind speed, 30s average power, generator speed, generator winding temperature, generator drive end bearing temperature, generator non-drive end bearing temperature, nacelle temperature, and ambient temperature.
8. An electronic device, comprising a processor and a memory; characterized in that, The processor is connected to the memory; the memory is used to store executable program code; the processor runs the program corresponding to the executable program code by reading the executable program code stored in the memory, so as to perform the steps of the wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the wind turbine condition monitoring method based on multi-condition identification and spatiotemporal graph network as described in any one of claims 1-7.