Wind turbine temperature monitoring method based on graph space-time attention network

Abstract

The application discloses a kind of wind-driven generator temperature monitoring methods based on graph space-time attention network, comprising the following steps: the SCADA data collected are preprocessed, obtain dataset;With each sensor as node, the top-k neighborhood relationship of each node is calculated, the edge weight between nodes is calculated by Gaussian kernel function, the weighted adjacency matrix is constructed based on the neighborhood relationship and edge weight of all nodes, so that the wind-driven generator multi-sensor time series graph is obtained, the time series data collected at different time stamps is assigned to node time series feature, form space-time graph structure, and output each sensor time series feature;Graph attention network is introduced into wind-driven generator temperature state monitoring, modeling the globality and connectivity features of multi-sensor network.The application significantly improves the generator temperature prediction accuracy by fusing graph space-time dual-dimensional attention mechanism and key variable screening: redundant variables are removed by using the maximum information coefficient MIC, and noise interference is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power, in particular to a wind turbine temperature monitoring method based on a graph spatio-temporal attention network. BACKGROUND

[0002] As the core equipment of wind energy conversion into electrical energy, the safe and stable operation of wind turbine is of great importance to energy supply efficiency and system reliability. As the key power component of wind turbine, the temperature state of wind turbine directly reflects the equipment operating condition. If the temperature anomaly is not monitored and warned in time, it may lead to unit shutdown, component damage and even safety accidents, causing huge economic losses. Therefore, it is of great significance to realize accurate prediction and early warning of wind turbine temperature for ensuring the safe operation of the unit and reducing maintenance costs.

[0003] Currently, wind turbine condition monitoring mainly relies on the multi-sensor data collected by the SCADA (Supervisory Control and Data Acquisition) system, including temperature, power, speed and other key parameters. However, SCADA data has the characteristics of multi-source heterogeneity and time series dynamic change, and there is a complex spatial correlation between nodes (sensors) in the multi-sensor network. Traditional monitoring methods are difficult to effectively model such global and connectivity characteristics.

[0004] Existing prediction methods based on SCADA data mainly focus on single sensor time series feature extraction, such as using LSTM and other time series models to process time series, but ignoring the spatial correlation between sensors, resulting in insufficient use of multi-source information. Although some methods introduce graph models to model spatial relationships, they lack specific attention to key sensor nodes and important time series segments, making it difficult to capture weak change characteristics under dynamic operating conditions. In addition, there are often missing values and outliers (such as shutdown data and limited power data) in SCADA data. If not properly preprocessed, it will seriously affect the subsequent prediction accuracy. At the same time, if the variable selection process fails to effectively eliminate redundant information, it will also increase the complexity of the model and reduce its reliability.

[0005] With the development of graph neural networks and attention mechanisms, graph attention network (GAT) provides a new way to model multi-sensor spatial correlation, but its combination with time series models still lacks collaborative mining of global-local features, resulting in insufficient prediction accuracy and timeliness under complex operating conditions. Therefore, how to effectively handle SCADA data noise, fully integrate multi-sensor spatio-temporal features, and accurately capture the trend of generator temperature changes to achieve early warning has become a key problem to be solved in the field of wind turbine condition monitoring. SUMMARY

[0006] To solve the above technical problems, the present application provides a wind turbine temperature monitoring method based on a graph spatio-temporal attention network, which is simple in algorithm and high in monitoring accuracy.

[0007] The technical scheme for solving the above technical problems is: a wind turbine temperature monitoring method based on a graph spatio-temporal attention network, comprising the following steps:

[0008] Step one: pre-process the collected SCADA data to obtain a data set;

[0009] Step two: process the data set by Min-max normalization, adopt a top-k nearest neighbor algorithm, take each sensor as a node, calculate the top-k nearest neighbor relationship of each node, calculate the edge weight between nodes by a Gaussian kernel function, construct a weighted adjacency matrix based on the nearest neighbor relationship and edge weight of all nodes, thereby obtaining a wind turbine multi-sensor time series graph, and the wind turbine multi-sensor time series graph is denoted as tKNNG, tKNNG is formed by assigning the time series data collected at different time stamps to the node time series features, forming a spatio-temporal graph structure, and outputting the time series features of each sensor;

[0010] Step three: introduce a graph attention network into the wind turbine temperature state monitoring, which is used to capture the global and connectivity features of the multi-sensor network: adopt a graph attention layer GAT to capture the internal structure information of the wind turbine, then extract the time series features of each sensor through a global-local attention embedding layer and an LSTM layer, fuse to generate spatio-temporal features under normal working conditions, and input the spatio-temporal features into a final prediction layer to estimate the monitoring variables.

[0011] The above wind turbine temperature monitoring method based on a graph spatio-temporal attention network, in step one, the pre-processing process is: first, delete all sequence data containing missing values; then, further remove abnormal data, including: wind turbine shutdown data with zero active power, limited power operation data and abnormal value data, to obtain pre-processed data.

[0012] The above wind turbine temperature monitoring method based on a graph spatio-temporal attention network, in step one, the pre-processed data is quantified by a maximum information coefficient MIC to calculate the correlation between feature variables, calculate the maximum information coefficient value between each sensor monitoring variable and the generator temperature state variable, and select the key variables highly related to the temperature state, and the key variables constitute a data set; the maximum information coefficient calculation formula is as follows:

[0013] ;

[0014] Wherein, is the maximum mutual information coefficient between variable A and variable B; represents the mutual information between variable and variable B; The normalized denominator represents the theoretical maximum value that mutual information can reach under the current grid size; and respectively along the variable and variables The number of grid cells in each direction; Representing variables The joint probability with variable B; and Variables The marginal probability density of variable B; , Representing variables respectively The integral over the bicontinuous domain of variables B and AB; This is an adjustable parameter.

[0015] In the above-mentioned wind turbine temperature monitoring method based on graph spatiotemporal attention network, in step two, the dataset is first processed by Min-max normalization:

[0016] ;

[0017] In the formula, Indicates the first The first sensor Time-series SCADA data, For the first The minimum value of sensor data, For the first The maximum value of each sensor data point These are the normalized data values.

[0018] In the above-mentioned wind turbine temperature monitoring method based on graph spatiotemporal attention networks, in step two, the first... Nodes The formal definition of the top-k nearest neighbor set is:

[0019] ;

[0020] in, Indicates the first Nodes The nearest neighbor set; function From dynamic sample sets The first one was selected from the middle Nodes The top-k nearest neighbors; Indicates the number of nearest neighbors. Represents a sliding window containing A subset of the number of samples express The first in One element;

[0021] In tKNNG, the edge weights between nodes are calculated using a Gaussian kernel weight function, defined as follows:

[0022] ;

[0023] in, For the first Nodes With the Nodes The right of the border; and Representing the first Nodes and the Nodes eigenvectors; express and The Euclidean distance between them; This represents the Gaussian kernel bandwidth parameter.

[0024] In the above-mentioned wind turbine temperature monitoring method based on graph spatiotemporal attention network, after calculating the nearest neighbor relationship and edge weight of each node in step two, a sparse matrix Z in coordinate format is first generated. The position of each non-zero element in Z corresponds to a connected pair of nodes, and the value of the non-zero element is the edge weight calculated by the Gaussian kernel function. Based on the sparse matrix Z, a weighted adjacency matrix is ​​derived. The weighted adjacency matrix fully represents the connection relationship and association strength between all nodes in tKNNG.

[0025] make Indicates time step The current state parameter frames are recorded in the form of a spatiotemporal graph structured data matrix. These parameter frames do not exist independently, but are interconnected through pairwise connections in the spatiotemporal graph. It is considered as a graph signal defined on G.

[0026] The above-described wind turbine temperature monitoring method based on graph spatiotemporal attention networks, in step three, utilizes historical state parameter data to predict future state parameters during wind turbine temperature state monitoring. Specifically, the historical state parameter data is described as a time series over a multi-sensor network of the wind turbine; the multi-sensor network is represented as... , ,in The node feature matrix, , Let n be the real number field, and n be the number of nodes, i.e., the number of sensors. is the feature length, i.e., the length of the input time series; E is the set of edges connecting the nodes; It is an adjacency matrix. , its first Line number Column elements Represents a node With nodes The correlation strength between them is calculated using the top-k nearest neighbor algorithm; the historical state information of the wind turbine is defined as a sequence on a multi-sensor network; therefore, the problem of predicting the temperature state parameters of the wind turbine is formulated as follows:

[0027] Given And historical state information, aiming to build a model, the model with a length of new sequence of As input, , for The Middle One element, predict the subsequent elements. State information at each time step , .

[0028] The above-mentioned wind turbine temperature monitoring method based on graph spatiotemporal attention networks, in step three, the process of capturing the internal structural information of the wind turbine using a graph attention layer (GAT) is as follows:

[0029] Using Graph Attention Layer (GAT), the first... Nodes Output characteristics The expression is as follows:

[0030] ;

[0031] ;

[0032] In the formula, Indicates the first The node is the first Normalized attention coefficients of each node; and They represent the first Nodes and the Nodes The new feature vector obtained by linear transformation of the original input feature vector; The symbol "" indicates that two vectors are concatenated; , Represents a new feature dimension for graph nodes; Represents the learnable parameters of a single-layer feedforward neural network; superscript Represents the transpose of a matrix; Indicates the first a node a set of neighborhood nodes of the node denotes a first neighbor node of the node a first neighbor node of the node a new feature vector after linear transformation of the feature vector of the first neighbor node of the node a new feature vector after linear transformation of the feature vector of the first neighbor node of the node M is the number of heads of the multi-head attention mechanism a normalized attention coefficient calculated by the i-th attention head ReLU represents the ReLU function; LeakReLU represents the LeakReLU activation function; in order to realize the integration of the spatial features of all nodes, the output of each node is fused through the following splicing operation:

[0033] ;

[0034] wherein, represents a spatial feature matrix modeled from G by GAT, .

[0035] The wind turbine temperature monitoring method based on the graph space-time attention network, in step three, the global-local attention embedding layer and the LSTM layer are used to extract the time sequence features of each sensor, and the process of fusing and generating the space-time features under normal working conditions is:

[0036] Since the output data of GAT is a graph sequence with a window length L, the global-local attention embedding layer uses a sliding window technology to calculate the contribution of each vector to the overall sequence. The calculation process of the global-local attention embedding layer is described as:

[0037] ;

[0038] ;

[0039] ;

[0040] wherein, represents the feature matrix of GAT at time step ; tanh represents the hyperbolic tangent function; is a learnable parameter matrix, , is the transpose of ; characterizes the local features of time step in the window; represents the local importance coefficient after normalization function processing of time step ; characterizes the global information of the entire window;

[0041] The methods for obtaining global and local information are as follows:

[0042]

[0043] In the formula, This represents the fused vector generated by concatenating local features and global information. ; The information is then fed into the LSTM unit for in-depth extraction of timing information, and the update process follows the standard LSTM gating mechanism.

[0044] After completing the spatiotemporal feature extraction, the time step Output for The spatiotemporal information integration characteristics, , For time step The corresponding graph structure data, where H is the hidden layer dimension; finally, the last LSTM unit is processed through a linear regression layer for output. .

[0045] In the above-mentioned wind turbine temperature monitoring method based on graph spatiotemporal attention networks, step three defines the root mean square error (RMSE) of the prediction residual as the monitoring index, and the time step... The Nodes root mean square error The calculation formula is:

[0046]

[0047] in, For the number of nodes, Indicates window size. and They represent the first Nodes Mid-time step The predicted and actual values;

[0048] In addition, the exponentially weighted moving average (EWMA) control chart is used to define the anomaly detection threshold, and the mathematical expression is:

[0049]

[0050] in, Indicates at time step The calculated exponentially weighted moving average; Indicates time step The root mean square error, Indicates the influence weight of historical data. ;

[0051] The monitoring threshold is calculated, and the time step is The monitoring threshold Defined by the EWMA control chart:

[0052]

[0053] Wherein, , The standard deviation and the mean of the test data set output residual respectively, K is the determination coefficient of the threshold value; at the same time, the mean absolute error MAE and the root mean square error RMSE are used to quantitatively evaluate the prediction performance.

[0054] The beneficial effects of the present application are:

[0055] 1、The present application significantly improves the generator temperature prediction accuracy by fusing the graph space-time dual attention mechanism and key variable screening; the maximum information coefficient MIC is used to remove redundant variables, and the sensor data strongly related to the temperature state is screened out, so as to reduce noise interference; the graph attention network GAT and the global-local attention embedding layer are combined to simultaneously capture the spatial topological correlation and time dynamic dependence of the multi-sensor network, fully excavate the space-time characteristics, and make the prediction result more in line with the actual working condition.

[0056] 2、The present application realizes early and accurate early warning of abnormal working conditions based on dynamic graph structure construction and residual analysis: the nonlinear relationship between sensors is modeled by tKNNG, which can adaptively capture the change of running state; the model output residual trained by health data is combined with the EWMA control chart threshold determination mechanism to alarm in real time when the root mean square error RMSE exceeds the threshold, which can trigger the early warning at the abnormal budding stage, and ensure the safe operation of the fan.

[0057] 3、The present application designs an end-to-end lightweight prediction framework, which has high efficiency and engineering robustness: Min-max normalization, sliding window sampling and LSTM time series modeling are used to reduce the computational complexity and adapt to large-scale SCADA data processing; by removing abnormal data such as shutdown / power limit, and using multi-head GAT to enhance the feature expression ability, the stability and generalization ability of the model under complex working conditions are ensured, which provides reliable technical support for fan operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The flowchart of the present application.

[0059] Figure 2 The node MIC correlation coefficient heat map.

[0060] Figure 3 The top-k nearest neighbor generator multi-temperature sensor time series structure diagram constructed according to the experiment.

[0061] Figure 4 The diagram shows a model for predicting and warning the temperature of a wind turbine generator using a spatiotemporal attention network.

[0062] Figure 5 This is a temperature prediction diagram of the generator stator winding according to the method proposed in this invention.

[0063] Figure 6 This is a graph showing the generator stator winding temperature prediction using the STGNN method.

[0064] Figure 7 The graph shows the results of generator bearing temperature monitoring using the method proposed in this invention.

[0065] Figure 8 The graph shows the results of generator bearing temperature monitoring using the STGNN method.

[0066] Figure 9 A comparison chart of the changes in MAE (Magnitude of Effect) for different models.

[0067] Figure 10 A comparison chart showing the changes in RMSE for different model indicators. Detailed Implementation

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] like Figure 1 As shown, a method for monitoring the temperature of a wind turbine based on a graph spatiotemporal attention network includes the following steps:

[0070] Step 1: Preprocess the collected SCADA data to obtain a dataset.

[0071] The preprocessing process is as follows: First, all sequence data containing missing values ​​are deleted to avoid the impact of missing values ​​on the analysis; then, outlier data is further removed, including: wind turbine shutdown data with zero active power, power-limited operation data, and outlier data, to obtain the preprocessed data.

[0072] The SCADA system of a wind turbine records multi-dimensional variables. When an anomaly occurs in the turbine's operation, directly inputting all variables into a deep learning model will drastically increase the anomaly detection time and significantly reduce reliability. Therefore, it is necessary to select key variables that characterize the health status of the wind turbine. An enhanced method using the maximum information coefficient (MIC) variable selection technique is employed on the preprocessed data to identify SCADA data from various generator sensors that have a high impact on temperature status, thereby improving the reliability of the SCADA data.

[0073] like Figure 2As shown, the maximum information coefficient between the correlation of the characteristic variables is used to calculate the maximum information coefficient value between the monitoring variables of each sensor and the temperature state variable of the generator, and the key variables highly related to the temperature state are selected, and the key variables constitute the data set; Figure 2 The MIC index distribution is shown in the middle. Analysis shows that some variables (such as generator stator winding temperatures U1, U2, V1, V2, W1, and W2) have very high correlation coefficients; and some parameters have lower MIC values than other variables, indicating that they have weaker representation of the health status of the wind turbine generator, and should be removed to focus on key information. Accordingly, the experiment selects variables with a correlation coefficient of more than 0.6 to construct a graph structure data, and therefore selects 8 most relevant sensors. The maximum information coefficient calculation formula is as follows:

[0074] ;

[0075] wherein, is the maximum mutual information coefficient between variable A and variable B, and is a value between , the closer to 1, the stronger the relationship, and the closer to 0, the weaker the relationship; represents the mutual information between variable and variable B, which measures the degree of mutual dependence between variables; is the normalized denominator, representing the theoretical maximum value of mutual information under the current grid size; and are the number of grids divided along the variable and the variable direction respectively; represents the joint probability of variable and variable B; and are the marginal probability densities of variable and variable B respectively; , represent the integrals on the two continuous variable domains composed of variable and variable B respectively; is an adjustable parameter.

[0076] Step two: normalize the data set by Min-max normalization, use top-k neighbor algorithm, take each sensor as a node, calculate the top-k neighbor relationship of each node, and calculate the edge weight between nodes by Gaussian kernel function. Based on the neighbor relationship and edge weight of all nodes, a weighted adjacency matrix is constructed, and a wind turbine multi-sensor time series graph is obtained, which is denoted as tKNNG. tKNNG assigns time series data collected at different time stamps to node time series features to form a spatio-temporal graph structure, and outputs each sensor time series feature.

[0077] The data set is first processed by Min-max normalization:

[0078] ;

[0079] wherein, denotes the first time series SCADA data of the th sensor, is the minimum value of the th sensor data, is the maximum value of the th sensor data, is the normalized data value.

[0080] As shown in FIG. 1, a top-k neighbor algorithm is adopted to calculate the top-k neighbor relationship of each node, with each sensor as a node, and the edge weight between nodes is calculated by a Gaussian kernel function. A weighted adjacency matrix is constructed based on the neighbor relationship and edge weight of all nodes, so as to obtain a wind turbine multi-sensor time series graph. Specifically, the top-k neighbor set of the Figure 3 th node is defined as:

[0081] ;

[0082] wherein, denotes the neighbor set of the th node ; the function filters the top-k nearest neighbors of the th node from the dynamic sample set ; denotes the number of neighbors, represents a subset containing samples in a sliding window, denotes the th element in ;

[0083] In the tKNNG, the edge weight between nodes is calculated by a Gaussian kernel weight function, which is defined as follows:

[0084] ;

[0085] wherein, is the edge weight of the th node and the th node ; and respectively represent the​ a node and the a node characteristic vector; denotes and the Euclidean distance between two nodes, the value of the distance measures the absolute difference between two nodes in the feature space, and the distance of 0 represents complete similarity, and the larger the distance represents the greater the difference; denotes the Gaussian kernel bandwidth parameter.

[0086] After calculating the near-neighbor relationship and edge weight of each node, first, a sparse matrix Z in coordinate format is generated, the position of each non-zero element in Z corresponds to the existing connected node pair, and the value of the non-zero element is the edge weight calculated by the Gaussian kernel function. Based on the sparse matrix Z, a weighted adjacency matrix is derived, which fully characterizes the connection relationship and correlation strength between all nodes in the tKNNG;

[0087] First, each sensor after normalization is directly regarded as a node. Then, the selected variables are sequentially numbered from 1 to 8 (in the experiment, in order: generator bearing A, B temperature, generator stator winding temperature U1, U2, V1, V2, W1, W2). Then, the tKNNG construction method is applied to determine the neighbors of each node, and the constructed wind turbine multi-sensor time series diagram is shown in Figure 3 From Figure 3 , it can be seen that when is set to 5, 5 neighbors are found for each node.

[0088] After the tKNNG is constructed, the time series data collected at different time stamps are assigned to the node features, and a space-time graph structure is formed, denotes the current state parameter frame at time step , which is recorded in the form of a space-time graph structure data matrix. These parameter frames do not exist independently, but are related to each other through the pair-wise connection in the space-time graph, and therefore is regarded as a graph signal defined on G .

[0089] Step three: Introducing the graph attention network into the wind turbine temperature state monitoring for capturing the global and connectivity features of the multi-sensor network: using the graph attention layer GAT to capture the internal structure information of the wind turbine, and then extracting the time series features of each sensor through the global-local attention embedding layer and the LSTM layer, and fusing to generate the space-time features under normal working conditions. The space-time features are input into the final prediction layer to estimate the monitoring variables.

[0090] In wind turbine temperature state monitoring, historical state parameter data is used to predict future state parameters. Specifically, the historical state parameter data is described as a time series on a wind turbine multi-sensor network; the multi-sensor network is represented as , where is a node feature matrix, , is a real field, and n is the number of nodes, i.e., the number of sensors; is the feature length, i.e., the input time series length; E is a set of edges connecting nodes; is an adjacency matrix, whose th column element represents the association strength between node and node , which is calculated by a top-k neighbor algorithm; the historical state information of the wind turbine is defined as a sequence on the multi-sensor network; therefore, the wind turbine temperature state parameter prediction problem is formulated as:

[0091] Given and historical state information, the goal is to build a model that takes a new sequence of length as input, , is the th element in , and predicts the state information , at the next time steps.

[0092] The temperature sensors carried by the wind turbine generator are not independent of each other, and the information features of each temperature sensor are coupled with each other, and each parameter is affected by other components of the generator. The correlation between state parameters at different positions of the generator is regarded as an inherent spatial feature, and tKNNG is a mathematical encoding of this sensor correlation. Sensors that are connected are more likely to share similar features. Specifically, in the prediction of generator temperature state parameters, the greater the edge weight between two nodes, the more likely it is that the sensor data will interact. To capture the spatial dependency relationship from the tKNNG topology, as shown in Figure 4 , a graph attention network (GAT) is used to learn the structure-aware node features. The process of using the graph attention layer GAT to capture the internal structure information of the wind turbine is as follows:

[0093] Using the graph attention layer GAT, the output feature of the th node is expressed as:

[0094] ;

[0095] ;

[0096] In the formula, Indicates the first The node is the first Normalized attention coefficients of each node; and They represent the first Nodes and the Nodes The new feature vector obtained by linear transformation of the original input feature vector; The symbol "" indicates that two vectors are concatenated; , Represents a new feature dimension for graph nodes; Represents the learnable parameters of a single-layer feedforward neural network; superscript Represents the transpose of a matrix; Indicates the first Nodes The set of neighboring nodes; express The One neighboring node; express The The new feature vector obtained by linearly transforming the feature vectors of the neighboring nodes; M is the number of heads in the multi-head attention mechanism; Indicates the first Normalized attention coefficients calculated by each attention head; ReLU represents the ReLU function; LeakReLU represents the LeakReLU activation function;

[0097] To achieve the integration of spatial features from all nodes, the outputs of each node are concatenated to achieve feature fusion:

[0098] ;

[0099] In the formula, This represents the spatial feature matrix obtained by modeling from G using GAT. .

[0100] After completing spatial dependency modeling, such as Figure 4 As shown, spatiotemporal features are collaboratively integrated through a global-local attention embedding layer and an LSTM layer. Given that the GAT output data is a graph sequence with a window length L, the global-local attention embedding layer uses a sliding window technique to calculate the contribution of each vector to the overall sequence. The calculation process of the global-local attention embedding layer is described as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] wherein, represents the feature matrix of GAT at time step ; ; tanh represents the hyperbolic tangent function; is a learnable parameter matrix, , is the transpose of ; characterizes the local feature of time step within the window; represents the local importance coefficient of time step after being processed by a normalization function; characterizes the global information of the entire window;

[0105] The global and local information is obtained as follows:

[0106]

[0107] wherein, represents the fusion vector generated by splicing the local feature and the global information, ; Then, the input is input into an LSTM unit for deep extraction of time sequence information, and the updating process follows the standard LSTM gating mechanism.

[0108] After the spatio-temporal feature extraction is completed, the output of time step is the spatio-temporal information integrated feature of , , is the graph structure data corresponding to time step , H is the hidden layer dimension; finally, the output of the last LSTM unit is processed through a linear regression layer .

[0109] As shown in Figure 5 , the state prediction of node 3 (i.e., the temperature of the generator stator winding U1) in G for one day is shown.

[0110] The fan operating state is complex and variable, and an evaluation threshold needs to be set to determine the working condition. There is always a residual error between the actual value and the predicted value in the prediction model, but the graph temporal-spatial attention network is trained based on the health running period data, and the model parameters have fitted the trend of the health data. In the state monitoring process, when the normal cycle data is input, the data and the model parameters are highly adapted, and the output residual error is small; on the contrary, when the abnormal cycle data is input, the data and the model parameters are mismatched, resulting in that the output residual error is significantly larger than the training error. Therefore, the processed prediction residual error can be used to evaluate the health state of the wind turbine generator. The root mean square error RMSE of the prediction residual error is defined as a monitoring index, and the time step of the th node of the th node is calculated as follows:

[0111]

[0112] wherein, is the number of nodes, denotes the window size, and denote the predicted value and the actual value of the time step in the th node ;

[0113] In addition, the exponential weighted moving average EWMA control chart is used to define the abnormality judgment threshold, and the mathematical expression is as follows:

[0114]

[0115] wherein, denotes the exponential weighted moving average value calculated at the time step ; denotes the mean value of the root mean square error at the time step , denotes the influence weight of the historical data, ;

[0116] The monitoring threshold is calculated, and the monitoring threshold at the time step is defined by the EWMA control chart as follows:

[0117]

[0118] wherein, , denote the standard deviation and the mean value of the output residual error of the test data set respectively, and K is the determination coefficient of the threshold.

[0119] As Figure 7As shown in FIG. 6, the temperature monitoring result of the node 1 (i.e., the temperature of the generator bearing A) in G is shown. At the same time, the mean absolute error MAE and the root mean square error RMSE are used to quantitatively evaluate the prediction performance of the method.

[0120] In order to verify the effectiveness of the method proposed in the application, the method proposed in the application is compared with other methods, and the comparison method is set as follows:

[0121] (1) Long short-term memory network LSTM: directly using the original temperature time series data of multiple measuring points of the generator as the input of the LSTM model;

[0122] (2) Gated recurrent unit GRU: directly using the original temperature time series data of multiple measuring points of the generator as the input of the GRU model;

[0123] (3) Convolutional neural network-gated recurrent unit hybrid model CNN-GRU: regarding the temperature time series data of multiple measuring points of the generator as multi-channel input, first using CNN to extract local space-time features, and then inputting the feature sequence into GRU for modeling;

[0124] (4) Spatio-temporal graph neural network STGNN: constructing the temperature sensor nodes on the generator into a graph structure, and using a spatio-temporal graph convolutional neural network to model the time and space dependence of the temperature data at the same time;

[0125] At the same time, the prediction and monitoring results of the generator temperature of the wind turbine are given, and compared with the existing method. The prediction result of the generator temperature is as shown in FIG. 6. Figures 5-8 In the generator temperature prediction experiment, it can be seen that the coincidence degree of the prediction value curve and the actual value curve of STGNN has a relatively obvious deviation, the mean absolute error reaches 1.7535, the fluctuation range is large, and the root mean square error also reaches 2.3852, which shows that the prediction result is not accurate enough. In comparison, it can be clearly seen that the prediction value curve of the method proposed in the application is basically coincided with the actual value curve, and the mean absolute error is 0.7675, which shows that the method proposed in the application has effectiveness and correctness, and fully shows that the wind turbine temperature monitoring method based on the graph spatio-temporal attention network has achieved good results in the prediction of the generator temperature.

[0126]

[0127] As can be seen from Table 1, by comparing the performance of different models in the prediction of the generator temperature, the MAE and RMSE of the method proposed in the application are the smallest except for the node 1 (the temperature of the generator bearing A), which are smaller than those of other methods. Overall, compared with other methods, the proposed method has achieved remarkable results, and the prediction performance is the best. The prediction index visualization diagram is as shown in FIG. 6.Figure 9 、 Figure 10 As shown in Figs. 6 and 7, the advantages and disadvantages of various models can be more intuitively reflected. In summary, the wind turbine generator temperature monitoring method based on the graph spatio-temporal attention network can well predict the health period state trend of the wind turbine generator and can process complex data changes.

[0128] To verify the effectiveness of the wind turbine generator temperature state monitoring method, the method is compared with the existing new method, and SCADA data of the generator with faults in a day are used for comparison. Figure 7 and Figure 8 The abnormal monitoring effect of the generator bearing temperature is shown, wherein the root mean square error RMSE of the temperature prediction residual is defined as the monitoring index, and the upper control limit UCL is calculated by using the exponential weighted moving average EWMA control chart. Figure 7 and Figure 8 The monitoring results in Figs. 8 and 9 show that the STGNN method is not sensitive to abnormal temperature information, and the abnormality of the generator bearing temperature is not detected at many time points with abnormal values, and the RMSE is still continuously lower than the upper control limit when the fault occurs, indicating that the STGNN method has the risk of missing the alarm and cannot timely detect the early abnormal operation state of the generator. The method successfully detects the abnormal change of the generator bearing temperature and successfully issues an abnormal alarm at multiple time points, timely identifies weak fault information and successfully detects the abnormality, realizes accurate and reliable abnormal alarm, and fully verifies the superiority of the method in the wind turbine generator temperature abnormality detection.

[0129] In summary, the method fuses the dynamic graph structure and the spatio-temporal attention mechanism, proposes a wind turbine generator temperature monitoring method based on the graph spatio-temporal attention network, realizes high-precision temperature prediction and early abnormality diagnosis through multi-sensor data collaborative modeling and residual early warning mechanism. The four-stage innovation of MIC-driven key sensor screening, tKNNG dynamic graph construction, spatio-temporal attention fusion and residual EWMA early warning solves the problems of insufficient precision and abnormality detection lag in the traditional method in the generator temperature prediction. The experiment shows that the MAE is more than 60% better than the benchmark model in the prediction task, which can better predict the future trend change of the generator temperature; in the state monitoring, the weak fault information is timely identified and the abnormality is successfully detected, early fault alarm is realized, and high-reliability technical support is provided for intelligent operation and maintenance of the wind turbine.​

Claims

1. A method for monitoring the temperature of a wind turbine generator based on a graph spatiotemporal attention network, characterized in that, Includes the following steps: Step 1: Preprocess the collected SCADA data to obtain a dataset; Step 2: The dataset is processed by Min-max normalization. The top-k nearest neighbor algorithm is used to calculate the top-k nearest neighbor relationship of each node, with each sensor as a node. The edge weights between nodes are calculated using a Gaussian kernel function. A weighted adjacency matrix is ​​constructed based on the nearest neighbor relationships and edge weights of all nodes, thereby obtaining the multi-sensor time series graph of the wind turbine. The multi-sensor time series graph of the wind turbine is denoted as tKNNG. tKNNG assigns the time series features of the nodes to the time series data collected at different timestamps, forming a spatiotemporal graph structure and outputting the time series features of each sensor. Step 3: Introduce graph attention network into wind turbine temperature status monitoring to capture the global and connectivity features of multi-sensor network: Use graph attention layer GAT to capture the internal structure information of wind turbine, then extract the temporal features of each sensor through global-local attention embedding layer and LSTM layer, fuse them to generate spatiotemporal features under normal operating conditions, and input the spatiotemporal features into the final prediction layer to estimate the monitoring variables. In wind turbine temperature condition monitoring, historical state parameter data is used to predict future state parameters. Specifically, the historical state parameter data is described as a time series over a multi-sensor network of the wind turbine; the multi-sensor network is represented as... , ,in The node feature matrix, , Let n be the real number field, and n be the number of nodes, i.e., the number of sensors. is the feature length, i.e., the length of the input time series; E is the set of edges connecting the nodes; It is an adjacency matrix. , its first Line number Column elements Represents a node With nodes The correlation strength between them is calculated using the top-k nearest neighbor algorithm; the historical state information of the wind turbine is defined as a sequence on a multi-sensor network; therefore, the problem of predicting the temperature state parameters of the wind turbine is formulated as follows: Given And historical state information, aiming to build a model, the model with a length of new sequence of As input, , for The Middle One element, predict the subsequent elements. State information at each time step , ; The process of capturing internal structural information of a wind turbine using Graph Attention Layer (GAT) is as follows: Using Graph Attention Layer (GAT), the first... Nodes Output characteristics The expression is as follows: ； ； In the formula, Indicates the first The node is the first Normalized attention coefficients of each node; and They represent the first Nodes and the Nodes The new feature vector obtained by linear transformation of the original input feature vector; The symbol "" indicates that two vectors are concatenated; , Represents a new feature dimension for graph nodes; Represents the learnable parameters of a single-layer feedforward neural network; superscript Represents the transpose of a matrix; Indicates the first Nodes The set of neighboring nodes; express The One neighboring node; express The The new feature vector obtained by linearly transforming the feature vectors of the neighboring nodes; M is the number of heads in the multi-head attention mechanism; Indicates the first Normalized attention coefficients calculated by each attention head; ReLU represents the ReLU function; LeakReLU represents the LeakReLU activation function; to achieve the integration of spatial features of all nodes, the outputs of each node are concatenated as follows to achieve feature fusion: ； In the formula, This represents the spatial feature matrix obtained by modeling from G using GAT. ; The process of extracting temporal features from each sensor using a global-local attention embedding layer and an LSTM layer, and then fusing them to generate spatiotemporal features under normal operating conditions, is as follows: Given that the output data of GAT is a sequence of graphs with a window length of L, the global-local attention embedding layer uses a sliding window technique to calculate the contribution of each vector to the overall sequence. The calculation process of the global-local attention embedding layer is described as follows: ； ； ； In the formula, This indicates that GAT is at time step eigenmatrix ; tanh represents the hyperbolic tangent function; For learnable parameter matrix, , for transpose; Characterizing time steps within the window Local features; Indicates time step The local importance coefficient after processing with the normalization function; Represents global information for the entire window; The methods for obtaining global and local information are as follows: ； In the formula, This represents the fused vector generated by concatenating local features and global information. ; The information is then fed into the LSTM unit for in-depth extraction of timing information, and the update process follows the standard LSTM gating mechanism. After completing the spatiotemporal feature extraction, the time step Output for The spatiotemporal information integration characteristics, , For time steps The corresponding graph structure data, where H is the hidden layer dimension; finally, the last LSTM unit is processed through a linear regression layer for output. .

2. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 1, characterized in that, In step one, the preprocessing process is as follows: First, all sequence data containing missing values ​​are deleted; Subsequently, abnormal data was further removed, including wind turbine shutdown data with zero active power, power-limited operation data, and outlier data, resulting in preprocessed data.

3. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 1, characterized in that, In step one, the correlation between feature variables is quantified using the maximum information coefficient (MIC) on the preprocessed data. The maximum information coefficient value between each sensor-monitored variable and the generator temperature state variable is calculated, and key variables highly correlated with the temperature state are selected. These key variables constitute the dataset. The formula for calculating the maximum information coefficient is as follows: ； in, For variables A The maximum mutual information coefficient between variable B and variable C; Representing variables Mutual information between variable B and variable B; The normalized denominator represents the theoretical maximum value that mutual information can reach under the current grid size; and respectively along the variable and variables The number of grid cells in each direction; Representing variables The joint probability with variable B; and Variables The marginal probability density of variable B; , Representing variables respectively The integral over the bicontinuous domain of variables B and AB; This is an adjustable parameter.

4. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 3, characterized in that, In step two, the dataset is first processed by Min-max normalization: ； In the formula, Indicates the first The first sensor Time-series SCADA data; For the first The minimum value of sensor data; For the first The maximum value of data from each sensor; These are the normalized data values.

5. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 3, characterized in that, In step two, the first Nodes The formal definition of the top-k nearest neighbor set is: ； in, Indicates the first Nodes The nearest neighbor set; function From dynamic sample sets The first one was selected from the middle Nodes The top-k nearest neighbors; Indicates the number of nearest neighbors. Represents a sliding window containing A subset of the number of samples express The first in One element; In tKNNG, the edge weights between nodes are calculated using a Gaussian kernel weight function, defined as follows: ； in, For the first Nodes With the Nodes The right to the side; and Representing the first Nodes and the Nodes eigenvectors; express and The Euclidean distance between them; This represents the Gaussian kernel bandwidth parameter.

6. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 5, characterized in that, In step two, after calculating the nearest neighbor relationship and edge weight of each node, a sparse matrix Z in coordinate format is first generated. The position of each non-zero element in Z corresponds to a connected pair of nodes. The value of the non-zero element is the edge weight calculated by the Gaussian kernel function. Based on the sparse matrix Z, the weighted adjacency matrix is ​​derived. The weighted adjacency matrix fully represents the connection relationship and association strength between all nodes in tKNNG. make Indicates time step The current state parameter frames are recorded in the form of a spatiotemporal graph structured data matrix. These parameter frames do not exist independently, but are interconnected through pairwise connections in the spatiotemporal graph. It is considered as a graph signal defined on G.

7. The wind turbine temperature monitoring method based on graph spatiotemporal attention network according to claim 6, characterized in that, In step three, the root mean square error (RMSE) of the predicted residuals is defined as the monitoring index, and the time step... The Nodes root mean square error The calculation formula is: ； in, For the number of nodes, Indicates window size. and They represent the first Nodes Mid-time step The predicted and actual values; In addition, the exponentially weighted moving average (EWMA) control chart is used to define the anomaly detection threshold, and the mathematical expression is: ； in, Indicates at time step The calculated exponentially weighted moving average; Indicates time step The root mean square error, Indicates the influence weight of historical data. ; Calculate the monitoring threshold, time step Monitoring threshold Defined by EWMA control chart: ； in, , represents the standard deviation and mean of the output residuals of the test dataset, respectively, and K is the coefficient of determination for the threshold; mean absolute error (MAE) and root mean square error (RMSE) are used to quantitatively evaluate the prediction performance.

Images