Wind turbine generator state monitoring method based on multi-working-condition identification and ESN

Abstract

The invention discloses a wind turbine generator set state monitoring method based on multi-working-condition identification and an ESN. The method comprises the following steps: step 1, collecting operation data in a wind turbine generator set SCADA system; 2, carrying out feature engineering processing on the SCADA data; 3, performing operation condition division on the unit health data set by adopting a K-means clustering algorithm; 4, optimizing the echo state network model based on a differential evolution algorithm, establishing power prediction normal behavior models under different working conditions, and determining state monitoring thresholds of the unit under different working conditions in combination with power prediction residual analysis; and step 5, in an online monitoring stage, firstly carrying out feature engineering processing on real-time operation data of the wind turbine generator, then determining a current working condition by utilizing a working condition identification model, finally predicting active power through a corresponding echo state network model, and calculating a power residual error to judge whether the wind turbine generator runs normally or not. Practical application shows that the abnormal state of the unit can be found earlier than that of a traditional SCADA system when a fault occurs.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent state monitoring of wind turbines, and in particular to a wind turbine state monitoring method based on multi-operating condition identification and ESN. Background Art

[0002] In the face of the increasingly urgent energy crisis and to promote the sustainable development of the new energy field, wind turbines (abbreviated as wind turbines) play an important role as large and complex equipment that converts wind energy into electrical energy. Affected by the combined effects of the external environment and the complex electromechanical characteristics inside the unit, the operating state of the wind turbine is highly variable and uncertain, which can easily lead to unit failure. In addition, wind turbines are usually installed in areas with abundant wind energy and harsh natural environment, which undoubtedly increases the difficulty and cost of unit operation and maintenance. According to statistics, the operation and maintenance costs of large wind farms are as high as 15% to 35% of their total revenue. In order to effectively reduce the operation and maintenance costs and improve the operating stability and reliability of wind turbines, it is of great significance and engineering value to actively carry out intelligent operation and maintenance work such as condition monitoring.

[0003] With the development of artificial intelligence technology, deep learning methods have been widely used in wind turbine condition monitoring. Among the many deep learning algorithms, recurrent neural network (RNN) has excellent performance in time series data analysis due to its unique recurrent connection architecture and is often used for wind speed, power and other predictions. However, RNN has problems such as gradient vanishing and gradient explosion during training. In order to solve the problems of slow RNN training speed and memory disappearance, the echo state network (ESN) with powerful nonlinear time series modeling capabilities and efficient training methods is introduced into the field of wind turbine condition monitoring to improve the accuracy and effectiveness of wind turbine condition monitoring. During the service of wind turbines, the influence of environmental factors makes its operating state complex and changeable, and the operating parameters of the unit vary greatly under different operating conditions. At present, most wind turbine condition monitoring methods have not fully considered the operating state of the unit and only rely on a single monitoring model for condition monitoring, which to a certain extent limits the accuracy and effectiveness of condition monitoring.

[0004] Aiming at the complex and changeable operating conditions of wind turbines and the problems of gradient vanishing and gradient explosion in recurrent neural networks (RNNs), the present invention proposes a wind turbine state monitoring method based on multi-condition identification and ESN. Summary of the invention

[0005] Aiming at the problem that the operating status of wind turbines is difficult to identify and the challenges faced by RNN in time series prediction, a wind turbine status monitoring method based on multi-condition identification and ESN is proposed.

[0006] In order to solve the above technical problems, the technical solution provided by the present invention is:

[0007] A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN comprises the following steps:

[0008] Step 1: Collect the unit operation data in the wind turbine SCADA system, i.e. SCADA data, including environmental and component characteristics such as wind speed, wind direction, active power and generator speed; (wind turbine operation data includes active power, generator speed and generator non-drive end bearing temperature; environmental parameters include wind speed, wind direction and ambient temperature)

[0009] Step 2: Perform feature engineering on the SCADA data, including data cleaning, feature selection, and normalization, to construct a healthy data set for normal operation of the unit, that is, a data set that meets the design requirements and operates under the operating specifications;

[0010] Step 3: Use the K-means clustering algorithm to divide the unit health data set into operating conditions, construct data sets corresponding to different operating conditions, and combine the unit control strategy and t-distributed stochastic neighbor embedding (t-SNE) algorithm to verify and analyze the rationality of the multi-condition results; (The health data set refers to the data set after data cleaning, feature selection and normalization in the normal operating state of the unit, reflecting the normal working state and performance indicators of the unit)

[0011] Step 4: Based on the differential evolution (DE) algorithm, the echo state network (ESN) model is optimized to establish the normal behavior model of power prediction under different operating conditions, and the state monitoring threshold of the unit under different operating conditions is determined by combining the power prediction residual analysis;

[0012] Step 5: In the online monitoring phase, the real-time operation data of the wind turbine is first processed by feature engineering, and then the current operating condition is determined using the operating condition identification model. Finally, the active power is predicted through the corresponding echo state network model, and the power residual is calculated to determine whether the unit is operating normally. When the operating data exceeds the monitoring threshold for three consecutive times, an early warning is issued.

[0013] Further, the process of step 2 is:

[0014] Step 21: directly delete the data other than missing, zero power, limited power, cut-in wind speed and cut-out wind speed in the SCADA data, and use the local anomaly factor algorithm to detect and delete the outlier data;

[0015] Step 22: Using the maximum mutual information coefficient (MIC) algorithm, select features with high correlation with the active power of the wind turbine generator set; high correlation is defined as: MIC value greater than 0.5.

[0016] Step 23: To eliminate the impact of data dimension, perform maximum and minimum normalization on the processed data.

[0017] In step 2, the maximum mutual information coefficient algorithm is used to calculate the MIC value between the active power and the SCADA feature, and the features with high correlation are selected as the model input. The formula for calculating the MIC value is as follows:

[0018]

[0019] Where: p(x) and p(y) represent probability density functions; p(x, y) represents the joint probability density function; I * (X; Y) represents the maximum mutual information value among all grid divisions; a, b represent the number of grids in the x and y directions; n represents the number of samples; k is set to 0.6 based on experience. MIC The range is [0,1]. The closer it is to 1, the higher the correlation, and vice versa.

[0020] In order to eliminate the impact of data dimension, the cleaned data is normalized to the maximum and minimum values. The calculation formula is as follows:

[0021] x'=(xx min ) / (xx max ) (3)

[0022] In the formula: x represents the original data; x' represents the normalized data; x min and x max Represents the minimum and maximum values ​​in the original data.

[0023] In view of the complex and changeable operating conditions of wind turbines, it is difficult for a single model to fully and accurately learn the operating characteristics of wind turbines. Therefore, it is necessary to divide the operating conditions of wind turbines to improve the accuracy of later state monitoring. Further, the process of step 3 is:

[0024] Step 31: Use the elbow rule to determine the optimal number of clusters K for the K-means clustering model;

[0025] Step 32: According to the determined K value, the multi-dimensional SCADA data is input into the K-means clustering model to identify and classify different operating conditions of the unit, and a condition identification model is obtained by training the K-means clustering model, thereby determining the operating state of the wind turbine unit;

[0026] Step 33: Combine the unit control strategy and t-SNE algorithm to analyze and visualize the clustering results to ensure the rationality and accuracy of the operating condition division.

[0027] In step 3, the elbow rule finds the optimal K value by calculating the sum of squared errors (SSE), which refers to the sum of the squares of the distances from each point to its nearest cluster center. SSE usually decreases as the K value increases, because more cluster centers mean that the points in each cluster are closer to its center. Before the inflection point, increasing the number of clusters will significantly reduce SSE, but after this point, the contribution of increasing the number of clusters to reducing SSE becomes less obvious. The calculation formula is as follows:

[0028]

[0029] Where: K represents the number of cluster centers; C i represents the number of samples in the i-th cluster; m i represents the cluster center of the i-th cluster; ||x j -m i || 2 Represents the Euclidean distance from each sample in the i-th cluster to the cluster center.

[0030] As further described, affected by environmental factors such as wind speed and wind direction, the operating conditions of wind turbines are complex and changeable. Doubly-fed wind turbines mainly use variable speed constant frequency power generation technology, which makes the speed of the generator change with the wind speed, but keeps the grid frequency constant through other control means. According to the control strategy of the doubly-fed wind turbine, the operation stage of the unit can be divided into four stages: constant speed 1, variable speed, constant speed 2 and constant power. Combining the unit control strategy and t-SNE algorithm, the clustering results are analyzed and visualized to ensure the rationality and accuracy of the working condition division.

[0031] Further, the process of step 4 is:

[0032] Step 41: According to the operating condition division model, the SCADA historical health data is divided into several data sets, and a wind turbine power prediction model based on an echo state network is established for the data sets under different operating conditions;

[0033] Step 42: using a differential evolution (DE) algorithm to optimize key parameters of the echo state network model under various working conditions, including the number of neurons in the reserve pool N, sparsity α, and spectrum radius ρ;

[0034] Step 43: The power residual between the predicted power and the actual operating power is used as the wind turbine monitoring index, and the state monitoring threshold of the wind turbine under multiple working conditions is set by using statistical methods.

[0035] In step 4, the echo state network (ESN) model includes an input layer, a dynamic reservoir and an output layer. The input layer is used to receive input signals to activate the network; the dynamic reservoir replaces the hidden layer of the traditional RNN, and the reservoir contains many randomly sparsely connected neurons for processing the signals from the input layer and the signals of the reservoir itself at the previous moment; the output layer is used to generate output signals. The prediction performance of the ESN model is mainly affected by the reservoir parameters, among which the key parameters include the number of reservoir neurons N, sparsity α and spectral radius ρ.

[0036] Assume that the input layer has M neurons, the reservoir has N neurons, and the output layer has L neurons. At time t, the input layer state, the reservoir state, and the output layer state are shown in formulas (5), (6), and (7), respectively.

[0037] u(t)=[u1(t),u2(t),...,u M (t)] T (5)

[0038] x(t)=[x1(t),x2(t),...x N (t)] T (6)

[0039] y(t)=[y1(t),y2(t),...y L (t)] T (7)

[0040] At time t+1, the state of the storage pool is updated according to formula (8), and the state of the output layer is updated according to formula (9). Among them, the connection weight from the input layer to the storage pool is W in , the connection weight inside the reserve pool is W, and the connection weight from the reserve pool to the output layer is W out .

[0041] x(t+1)=f(W in ·u(t+1)+W·x(t)) (8)

[0042] y(t+1)=g(W out ·[u(t+1);x(t+1);y(t)]) (9)

[0043] Where f and g are the neuron activation functions inside the reservoir and the output layer, respectively.

[0044] Further described, the DE-ESN parameter optimization process is as follows:

[0045] (1) Decoding scheme: The gene dimension D of each individual in the DE population is 3. The search space of N is set according to the number of training samples and the complexity of each unit operating condition. In order to reduce the scale of the search space and the computational complexity of the algorithm, the search space of N is discretized with an interval of 100; the search space of α is [0.01, 0.1]; the search space of ρ is [0.1, 0.99].

[0046] (2) Fitness function: For each individual in the DE population, the difference between the predicted value of the ESN model (t = 1, 2, ..., k; k represents the number of output samples) and the actual value is used as the fitness value. The root mean squared error (RMSE) is selected as the fitness function. The fitness value of the i-th individual is shown in formula (10).

[0047]

[0048] (3) Optimize the process; the specific steps are as follows:

[0049] A. Population initialization. Set the population size NP, gene dimension D, and gene value range [U min ,U max ], mutation operator F, crossover probability CR, prediction accuracy μ and maximum number of iterations m. After generating the initial population, set the number of iterations G = 0.

[0050] B. Determination of iterative termination conditions: If the global optimal value reaches the prediction accuracy μ or the number of iterations reaches the maximum number of iterations m, terminate the DE iteration and go to step G; otherwise, execute step C.

[0051] C. Perform mutation, crossover and selection to generate the next generation of individuals.

[0052] D. Repeat step C until the next generation of population is generated.

[0053] E. Calculate the fitness value of the next generation population, determine the current global optimal value and its corresponding global optimal individual. F. Set G = G + 1, and then return to step B.

[0054] G. Decode the optimal individuals in the DE population and map them into the three key parameter values ​​of ESN.

[0055] H. Use the training set to train ESN and obtain a trained network.

[0056] I. Input the test set for prediction.

[0057] Further, the process of step 5 is:

[0058] Step 51: input the real-time operation data of the wind turbine generator set into the power monitoring model, and determine it as abnormal operation data when the predicted power residual exceeds the monitoring threshold;

[0059] Step 52: According to the early warning conditions, determine whether the operating state of the wind turbine generator set is abnormal.

[0060] In the present invention, the idea of ​​the wind turbine state monitoring method based on multi-condition identification and ESN is: first, the maximum mutual information coefficient (MIC) is used for input feature selection. Secondly, the K-means clustering algorithm is used to divide the unit operating conditions, and the rationality of the multi-condition results is verified and analyzed in combination with the unit control strategy and t-SNE algorithm. Then, the differential evolution algorithm (DE) is used to optimize the echo state network model to enhance its adaptability to complex conditions, so as to carry out active power prediction of wind turbines under different conditions. Combined with the power prediction residual analysis, the corresponding health threshold is determined for judging the unit operating status.

[0061] The beneficial effects of the present invention are as follows: a method for monitoring the status of a wind turbine based on multi-condition identification and ESN is proposed. When identifying the operating conditions of a wind turbine, the method uses a K-means clustering algorithm to construct a wind turbine operating status identification model, which effectively solves the problem of difficulty in dividing the operating conditions of the unit. At the same time, in order to solve the problems of gradient vanishing and gradient explosion faced during RNN training, an ESN model is introduced to predict the time series of wind turbines, and an ESN-based wind turbine operating status monitoring model under different operating conditions is constructed. Combined with the setting of dynamic thresholds, the problem that a single model is difficult to adapt to complex conditions is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is the overall framework diagram of the wind turbine condition monitoring method based on multi-condition identification and ESN;

[0063] Figure 2 This is a comparison chart of wind power data before and after cleaning;

[0064] Figure 3 It is the heat map of MIC values ​​between SCADA features;

[0065] Figure 4 It is the relationship between SSE and K value;

[0066] Figure 5 This is the clustering result diagram of wind turbine data;

[0067] Figure 6 It is the control strategy diagram of the wind turbine;

[0068] Figure 7 It is a schematic diagram of the relationship between wind speed, active power and main shaft speed in the actual operation data of wind turbines;

[0069] Figure 8 It is a scatter plot between wind speed, active power and main shaft speed of wind turbines under different working conditions;

[0070] Fig. 9 This is the dimensionality reduction visualization result diagram of wind turbine data;

[0071] Fig.10 It is the ESN network structure diagram;

[0072] Fig.11 This is the power prediction result diagram of the unit based on DE-ESN under different working conditions;

[0073] Fig.12 This is a monitoring result diagram of the operating status of the wind turbine. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings.

[0075] Reference Figures 1 to 12 A wind turbine condition monitoring method based on multi-condition identification and ESN is provided. The overall framework of the method is shown in the attached Figure 1 As shown, the following steps are included:

[0076] Step 1: Collect the unit operation data in the wind turbine SCADA system, i.e., SCADA data, including environmental and component characteristics such as wind speed, wind direction, active power, and generator speed;

[0077] Step 2: Perform feature engineering on the SCADA data, including data cleaning, feature selection, and normalization, to build a healthy data set for normal operation of the unit;

[0078] Step 21: Delete the missing, zero power, limited power, cut-in wind speed and cut-out wind speed data in the SCADA data directly, and use the local anomaly factor algorithm to detect and delete the outlier data. The effect before and after wind power data cleaning is shown in the attached figure. Figure 2 As shown;

[0079] Step 22: Use the maximum mutual information coefficient (MIC) algorithm to select features with high correlation with the active power of wind turbines. The active power of wind turbines is an important indicator reflecting their operating status. By calculating the MIC value between SCADA feature data and active power, the correlation between each feature and active power is analyzed, as shown in the attached figure. Figure 3 Table 1 lists the MIC values ​​between each SCADA feature and active power in detail, and finally selects the features with MIC values ​​greater than 0.5 as the input features of the model. This step ensures that the model can focus on the features that are closely related to active power and have a significant impact on the operating status, thereby improving the accuracy of the model prediction;

[0080] Table 1

[0081]

[0082] Step 23: To eliminate the impact of data dimension, perform maximum and minimum normalization on the processed data.

[0083] Step 3: Use the K-means clustering algorithm to divide the unit health data set into operating conditions, construct data sets corresponding to different operating conditions, and combine the unit control strategy and t-SNE algorithm to verify and analyze the rationality of the multi-condition results;

[0084] Step 31: Determine the optimal number of cluster centers K of the K-means clustering model through the elbow rule; the core evaluation index of the elbow method is the sum of squared errors (SSE). The multidimensional SCADA data selected in the feature engineering link is input into the K-means clustering model to identify and divide different operating conditions of the unit. By plotting the relationship between SSE and K value, it is determined that the optimal number of cluster centers K of the unit data is 4, and the clustering effect reaches an inflection point, as shown in the attached figure. Figure 4 As shown;

[0085] Step 32: Based on the determined K value, cluster analysis is performed on the historical healthy SCADA data to classify the wind turbine operating status. The wind turbine operating status is divided into four operating conditions. The distribution of data under different operating conditions is shown in the attached figure. Figure 5 The K-means clustering model is used to identify the operating conditions of wind turbines. On the one hand, it comprehensively considers the operating characteristics of wind turbines, and on the other hand, it provides effective support for the construction of subsequent multi-condition models and operating status monitoring;

[0086] Step 33: Combined with the unit control strategy and t-SNE algorithm, the clustering results are discussed, analyzed and visualized to ensure the rationality of the operating condition division. Specifically, affected by environmental factors such as wind speed and wind direction, the operating conditions of wind turbines are complex and changeable. Doubly fed wind turbines mainly use variable speed constant frequency power generation technology, which makes the speed of the generator change with the wind speed, but keeps the grid frequency constant through other control means. According to the control strategy of the doubly fed wind turbine, the operating stages of the unit can be divided into four stages: constant speed 1, variable speed, constant speed 2 and constant power, as shown in the attached figure. Figure 6 To further analyze the relationship between the K-means clustering algorithm results and the unit control strategy, the relationship between wind speed, active power and main shaft speed in the actual operation SCADA data after the working condition division is drawn, as shown in the attached figure. Figure 7It should be pointed out that due to the characteristics of wind turbines in terms of mechanical structure, electrical performance and control system design, the switching of unit control strategies caused by wind speed fluctuations usually requires a certain transition process rather than instantaneous completion. This leads to a certain overlap in the numerical ranges of characteristics between different operating conditions in the actual operation of the unit, resulting in deviations from the operating condition boundaries defined by the control strategy. Figure 8 The scatter plots between wind speed, active power and main shaft speed under different working conditions are shown. As can be seen from the figure, under working conditions 1, 3 and 4, although the main shaft speed fluctuates within a certain range, it remains relatively stable as a whole. The division of different working conditions is consistent with the unit control strategy, indicating that the analysis results of the K-means clustering algorithm are highly consistent with the unit control mechanism. In order to further explore the inherent structural characteristics of wind turbine SCADA data and simplify the data understanding and analysis process, the t-SNE algorithm is introduced. The algorithm can effectively map high-dimensional input features into low-dimensional space, and then clearly present the boundaries between different operating conditions on a two-dimensional plane, as shown in the attached figure. Fig. 9 This visualization result further verifies the rationality of the K-means clustering algorithm to divide the wind turbine operating data into four categories; step 33 is to explain the relationship between the K-means clustering result in step 32 and the unit control strategy. The four operating conditions are basically consistent with the four stages of unit operation.

[0087] Step 4: Based on the differential evolution algorithm (DE) to optimize the echo state network (ESN) model, establish the normal behavior model of power prediction under different operating conditions, and combine the power prediction residual analysis to determine the state monitoring threshold of the unit under different operating conditions;

[0088] Step 41: According to the operating condition division model, the SCADA historical health data is divided into several data sets, and a wind turbine power prediction model based on the echo state network is established for the data sets under different operating conditions. The ESN model is a new RNN improved model that solves the gradient disappearance and explosion problems existing in the traditional RNN model. The model contains an input layer, a dynamic reserve pool and an output layer. Its structure is shown in the attached figure. Fig.10 As shown;

[0089] Step 42: Use the differential evolution algorithm (DE) to optimize the key parameters of the echo state network model under various operating conditions, including the number of neurons in the reserve pool N, the sparsity α, and the spectral radius ρ. The randomly generated reserve pool has a significant impact on the prediction performance of the echo state network. Therefore, in order to obtain excellent prediction performance, it is particularly important to set appropriate values ​​for the key parameters of the reserve pool. Use the differential evolution algorithm (DE) to optimize these key parameters. Table 2 shows the optimal parameters of the DE-ESN model under different operating conditions of the wind turbine and its corresponding power prediction error. The prediction results of the active power of the wind turbine under different operating conditions of the unit are shown in the attached figure. Fig.11 As shown;

[0090] Step 43: The power residual between the predicted power and the actual operating power is used as the wind turbine monitoring index. The state monitoring threshold of the unit under multiple operating conditions is set by using statistical methods. The corresponding wind turbine state monitoring thresholds under different operating conditions are shown in Table 3.

[0091] Table 2

[0092]

[0093] Table 3

[0094]

[0095] Step 5: In the online monitoring phase, the real-time operation data of the wind turbine is first processed by feature engineering, and then the current operating condition is determined using the operating condition identification model. Finally, the active power is predicted through the corresponding echo state network model, and the power residual is calculated to determine whether the unit is operating normally. When the operating data exceeds the monitoring threshold for three consecutive times, an early warning is issued.

[0096] Step 51: input the real-time operation data of the wind turbine generator set into the state monitoring model, and determine it as abnormal operation data when the predicted power residual exceeds the monitoring threshold;

[0097] Step 52: According to the early warning conditions, determine whether the wind turbine operating state is abnormal. In order to verify the impact of operating condition division and dynamic threshold setting on the wind turbine state monitoring results, the proposed method is compared with the ESN state monitoring model without operating condition division. The results are shown in the attached figure. Fig.12 By comparison, it can be found that the division of working conditions can effectively reduce the false alarms of condition monitoring and can detect abnormal conditions of wind turbines earlier, thereby improving the accuracy of condition monitoring.

[0098] In summary, the wind turbine status monitoring method based on multi-condition identification and ESN proposed in the present invention can effectively identify different operating conditions of the unit and monitor its operating status. When a fault occurs, the abnormal state of the wind turbine can be discovered earlier than the SCADA system.

Claims

1. A wind turbine status monitoring method based on multi-operating condition identification and ESN, characterized in that: The steps include: Step 1: Collect the unit operation data in the wind turbine SCADA system, i.e., SCADA data, which includes wind turbine operation data and environmental parameters; the wind turbine operation data includes active power, generator speed, and generator non-drive end bearing temperature; the environmental parameters include wind speed, wind direction, and ambient temperature; Step 2: Perform feature engineering on the SCADA data, including data cleaning, feature selection, and normalization, to construct a data set in which the unit operates in accordance with design requirements and operating specifications, that is, a healthy data set under normal operating conditions; Step 3: Use the K-means clustering algorithm to divide the unit health data set into operating conditions and build an operating condition identification model; according to the results of different operating condition divisions, build the corresponding operating condition data set, and combine the unit control strategy and t-distributed random neighbor embedding t-SNE algorithm to verify and analyze the rationality of the multi-condition results; Step 4: Based on the differential evolution algorithm DE, the echo state network ESN model is optimized to establish a normal behavior model for power prediction under different working conditions, and combined with the power prediction residual analysis, the state monitoring threshold of the unit under different working conditions is determined; the power prediction normal behavior model is constructed through historical data under the normal operating state of the wind turbine unit, which is used to characterize the normal operating characteristics of the unit; Step 5: In the online monitoring phase, the real-time operating data of the wind turbine is first subjected to feature engineering processing, and then the operating condition recognition model is used to determine the current operating condition. Finally, the active power is predicted through the corresponding echo state network model, and the power residual is calculated to determine whether the unit is operating normally. When the operating data exceeds the monitoring threshold for three consecutive times, an early warning is issued.

2. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 1, characterized in that: The process of step 2 is: Step 21: directly delete the data other than missing, zero power, limited power, cut-in wind speed and cut-out wind speed in the SCADA data, and use the local anomaly factor algorithm to detect and delete the outlier data; Step 22: Using the maximum mutual information coefficient MIC algorithm, select the features whose MIC value with the active power of the wind turbine is greater than 0.5; Step 23: To eliminate the impact of data dimension, perform maximum and minimum normalization on the processed data.

3. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 2, characterized in that: The maximum mutual information coefficient algorithm is used to calculate the MIC value between active power and SCADA features. Features with MIC values ​​greater than 0.5 are selected as model inputs. The formula for calculating the MIC value is as follows: Where: p(x) and p(y) represent probability density functions; p(x, y) represents the joint probability density function; I * (X; Y) represents the maximum mutual information value in all grid divisions; a, b represent the number of grids in the x and y directions; n represents the number of samples; k is set to 0.6 based on experience; R MIC The range is [0,1]. The closer it is to 1, the higher the correlation is, and vice versa. In order to eliminate the impact of data dimension, the cleaned data is normalized to the maximum and minimum values. The calculation formula is as follows: x'=(x-x min ) / (x-x max ) (3) In the formula: x represents the original data; x' represents the normalized data; x min and x max Represents the minimum and maximum values ​​in the original data.

4. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 1, characterized in that: The process of step 3 is: Step 31: Use the elbow rule to determine the optimal number of clusters K for the K-means clustering model; Step 32: According to the determined K value, the multi-dimensional SCADA data is input into the K-means clustering model, different operating conditions of the unit are identified and divided, and an operating condition identification model is constructed to determine the operating status of the wind turbine unit; Step 33: According to the control strategy of the doubly fed wind turbine generator set, the operation stage of the unit can be divided into four stages: constant speed 1, variable speed, constant speed 2 and constant power; in order to analyze the relationship between the K-means clustering results and the control strategy, the relationship between the wind speed, active power and main shaft speed after the working condition division is used. At the same time, the t-SNE algorithm is used to map the high-dimensional input features to a two-dimensional plane, further verifying the rationality of the K-means clustering algorithm for the division of wind turbine operating data.

5. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 4, characterized in that: The elbow rule is to find the optimal K value by calculating the sum of squared errors (SSE). SSE refers to the sum of squares of the distances from each point to its nearest cluster center. The calculation formula is as follows: Where: K represents the number of cluster centers; C i represents the number of samples in the i-th cluster; m i represents the cluster center of the i-th cluster; ||x j -m i || 2 Represents the Euclidean distance from each sample in the i-th cluster to the cluster center.

6. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 1, characterized in that: The process of step 4 is: Step 41: According to the operating condition division model, the SCADA historical health data is divided into several data sets, and a wind turbine power prediction model based on the echo state network ESN is established for the data sets under different operating conditions; Step 42: using a differential evolution algorithm DE to optimize key parameters of the echo state network model under various working conditions, including the number of neurons in the reserve pool N, sparsity α, and spectrum radius ρ; Step 43: Use the power residual between the predicted power and the actual operating power as a monitoring indicator for the wind turbine set; and use a statistical method to set a state monitoring threshold for the set under multiple operating conditions.

7. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 6, characterized in that: The ESN model consists of an input layer, a dynamic reservoir, and an output layer. The input layer is used to receive input signals to activate the network. The dynamic reservoir replaces the hidden layer of the traditional RNN. The reservoir contains many randomly sparsely connected neurons, which are used to process the signals from the input layer and the signals of the reservoir itself at the previous moment. The output layer is used to generate output signals. Assume that the input layer has M neurons, the reservoir has N neurons, and the output layer has L neurons; at time t, the input layer state, the reservoir state, and the output layer state are shown in formulas (5), (6), and (7), respectively: u(t)=[u1(t),u2(t),...,u M (t)] T (5) x(t)=[x1(t),x2(t),...x N (t)] T (6) y(t)=[y1(t),y2(t),...y L (t)] T (7) At time t+1, the state of the storage pool is updated according to formula (8), and the state of the output layer is updated according to formula (9); where the connection weight from the input layer to the storage pool is W in , the connection weight inside the reserve pool is W, and the connection weight from the reserve pool to the output layer is W out ; x(t+1)=f(W in ·u(t+1)+W·x(t)) (8) y(t+1)=g(W out ·[u(t+1);x(t+1);y(t)]) (9) Where f and g are the neuron activation functions inside the reservoir and the output layer, respectively.

8. A method for monitoring the status of a wind turbine generator system based on multi-operating condition identification and ESN as claimed in claim 1, characterized in that: The process of step 5 is: Step 51: input the real-time operation data of the wind turbine generator set into the wind turbine generator set power prediction model, predict the active power of the generator set and calculate the power residual, and determine it as abnormal operation data when the predicted power residual exceeds the monitoring threshold; Step 52: According to the early warning conditions, determine whether the operating state of the wind turbine generator set is abnormal.

Images