Wind turbine generator and fault early warning method and device thereof, storage medium and program product

Abstract

The invention discloses a wind turbine generator and a fault early warning method and device thereof, a storage medium and a computer program product. The method comprises the following steps: processing historical data of the wind turbine generator to obtain a sample, and dividing the sample into a training set and a test set; after an improved BiGRU model is constructed based on CycleNet and a multi-head attention mechanism, the training set is used for training to obtain an operation prediction model of the wind turbine generator, a preset sliding window is added, the test set is used for testing to correct the operation prediction model of the wind turbine generator, and a fault prediction model of the wind turbine generator is obtained; and utilizing the fault prediction model of the wind turbine generator to predict whether the wind turbine generator has a fault based on the current data of the wind turbine generator so as to realize fault prediction of the wind turbine generator. According to the scheme, the fault early warning model is obtained based on CycleNet and multi-head attention mechanism learning, fault early warning of the variable pitch system is achieved under the complex weather condition, and the operation efficiency and reliability of wind power equipment are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and specifically relates to a fault warning method, device, wind turbine, storage medium and computer program product for a wind turbine set, and in particular to a fault warning method, device, wind turbine, storage medium and computer program product for a wind turbine set pitch system based on CycleNet and a multi-head attention mechanism. Background Art

[0002] The wind turbine pitch system is prone to failure or shutdown under complex and changeable weather conditions, such as extreme weather events such as strong winds, thunderstorms, and sandstorms, as well as factors such as high temperature, vibration, and wind speed changes. The fault diagnosis methods for wind turbines in related solutions mainly focus on vibration signal analysis and data-driven fault diagnosis methods. The lack of universality of the model makes it difficult to directly migrate and apply it to other systems of wind turbines or diversified fault warnings.

[0003] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0004] The purpose of the present invention is to provide a fault warning method, device, wind turbine, storage medium and computer program product for a wind turbine to solve the problems that the fault diagnosis methods for wind turbines in related schemes mainly focus on vibration signal analysis, data-driven fault diagnosis methods, etc., lack universality, and the accuracy of fault diagnosis of wind turbines under complex weather conditions cannot be guaranteed. The fault warning model is obtained by learning based on CycleNet and multi-head attention mechanism, and the fault warning of the wind turbine variable pitch system can be realized under complex weather conditions, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0005] The present invention provides a fault early warning method for a wind turbine generator set, comprising: collecting historical data of the wind turbine generator set from a SCADA system of the wind turbine generator set; processing the historical data of the wind turbine generator set to obtain samples, and dividing the samples into a training set and a test set; constructing a model based on CycleNet and a multi-head attention mechanism; using the training set to train the improved BiGRU model to obtain an improved BiGRU operation prediction model of the wind turbine generator set; after adding a preset sliding window to the operation prediction model of the wind turbine generator set, using the test set to perform a test to obtain a test result; and correcting the operation prediction model of the wind turbine generator set according to the test result to obtain a fault prediction model of the wind turbine generator set; acquiring current data of the wind turbine generator set; using the fault prediction model of the wind turbine generator set to predict whether the wind turbine generator set is faulty based on the current data of the wind turbine generator set, so as to realize fault early warning of the wind turbine generator set.

[0006] In some embodiments, the historical data of the wind turbine group is processed to obtain a sample, including: cleaning the historical data of the wind turbine group to obtain first processed data of the wind turbine group; using a Pearson correlation algorithm to extract characteristic data of a variable pitch system of the wind turbine group from the first processed data of the wind turbine group to obtain second processed data; after normalizing the second processed data, selecting operating data of the wind turbine group as a sample.

[0007] In some embodiments, the historical data of the wind turbine group is cleaned to obtain the first processed data of the wind turbine group, including: cleaning the abnormal data in the historical data of the wind turbine group according to a preset abnormal data range to obtain the first cleaned data; using the LOF algorithm to clean the discrete points and abnormal data in the first cleaned data again to obtain the first processed data of the wind turbine group.

[0008] In some embodiments, based on CycleNet and a multi-head attention mechanism, an improved BiGRU model is constructed, including: based on BiGRU, obtaining a forward GRU part and a reverse GRU part; using a CycleNet network, merging the forward GRU part and the reverse GRU part to obtain a merged model; after adding a dropout layer to the merged model, mapping the features in the merged model to an output space through a fully connected layer; in the output space, based on a multi-head attention mechanism, dividing the input of the output space into more than two heads, and calculating an attention matrix for each head; connecting the attention matrices of more than two heads in series to form a new matrix as the improved BiGRU model.

[0009] In some embodiments, after adding a preset sliding window to the operation prediction model of the wind turbine group, the test set is used to perform testing to obtain test results, including: after constructing the operation prediction model of the wind turbine group, using a sliding window statistics method, adding a preset sliding window, and performing fault warning on the data of the test set to obtain a warning result; according to the warning result, setting the upper and lower bounds of the residual threshold of the operation prediction model of the wind turbine group as an indicator for confirming whether an output result of the operation prediction model of the wind turbine group is a warning, and using the indicator as the test result.

[0010] In some embodiments, it also includes: after obtaining the operation prediction model of the wind turbine group, using the crown porcupine optimization algorithm to optimize the hyperparameters of the operation prediction model of the wind turbine group to obtain an improved BiGRU model based on CycleNet and a multi-head attention mechanism, recorded as an optimized CPO-BiGRU model; using the training set to train the optimized CPO-BiGRU model to obtain the optimized operation prediction model of the wind turbine group; after adding a preset sliding window to the optimized operation prediction model of the wind turbine group, using the test set to test to obtain a test result; and according to the test result, the optimized operation prediction model of the wind turbine group is corrected to obtain the optimized fault prediction model of the wind turbine group.

[0011] Matching the above method, the present invention provides a fault warning device for a wind turbine group on the other hand, comprising: an acquisition unit, configured to collect and obtain historical data of the wind turbine group from the SCADA system of the wind turbine group; a control unit, configured to process the historical data of the wind turbine group to obtain samples, and divide the samples into a training set and a test set; the control unit is also configured to construct a BiGRU model based on CycleNet and a multi-head attention mechanism; the control unit is also configured to train the improved BiGRU model using the training set to obtain an operation prediction model of the wind turbine group; the control unit is also configured to add a preset sliding window to the operation prediction model of the wind turbine group, and then use the test set to test to obtain a test result; and according to the test result, the operation prediction model of the wind turbine group is corrected to obtain a fault prediction model of the wind turbine group; the acquisition unit is also configured to obtain current data of the wind turbine group; the control unit is also configured to use the fault prediction model of the wind turbine group to predict whether the wind turbine group is faulty based on the current data of the wind turbine group, so as to realize fault warning of the wind turbine group.

[0012] Matching the above device, the present invention provides a wind turbine generator set on another aspect, including: the fault warning device of the wind turbine generator set mentioned above.

[0013] In accordance with the above method, the present invention provides a storage medium on another aspect, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the above-mentioned fault warning method for the wind turbine set.

[0014] In accordance with the above method, the present invention provides a computer program product on another aspect, including a computer program, which implements the steps of the above-mentioned fault warning method for wind turbines when executed by a processor.

[0015] Therefore, the scheme of the present invention extracts historical data of the wind turbine group, cleans, features and normalizes the historical data, selects wind turbine operation data in the wind turbine group as samples, and divides the samples into a training set and a test set; constructs a BiGRU model including a bidirectional gated recurrent unit and a multi-head attention mechanism, trains with the training set, and then tests and corrects with the test set to obtain a fault warning model; obtains the current data of the wind turbine group, so as to use the fault warning model to predict whether the wind turbine group is faulty based on the current data; thereby, the fault warning model is obtained by learning based on CycleNet and the multi-head attention mechanism, and the fault warning of the wind turbine variable pitch system can be realized under complex weather conditions, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0016] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a flow chart of an embodiment of a fault warning method for a wind turbine generator system of the present invention;

[0019] Figure 2 A schematic diagram of a flow chart of an embodiment of processing the historical data of the wind turbine generator set to obtain samples in the method of the present invention;

[0020] Figure 3 A schematic diagram of a flow chart of an embodiment of cleaning the historical data of the wind turbine generator set in the method of the present invention;

[0021] Figure 4 A flowchart of an embodiment of building an improved BiGRU model based on CycleNet and a multi-head attention mechanism in the method of the present invention;

[0022] Figure 5It is a flowchart of an embodiment of the method of the present invention that uses the test set to perform testing to obtain test results;

[0023] Figure 6 A schematic diagram of a flow chart of an embodiment of optimizing the operation prediction model of the wind turbine generator set in the method of the present invention;

[0024] Figure 7 It is a structural schematic diagram of an embodiment of a fault warning device for a wind turbine generator set of the present invention;

[0025] Figure 8 It is a schematic diagram of a fault warning process in a fault warning method for a wind turbine pitch system based on CycleNet and a multi-head attention mechanism of the present invention;

[0026] Fig. 9 It is a schematic diagram of the data processing flow in a fault warning method for a wind turbine pitch system based on CycleNet and a multi-head attention mechanism of the present invention;

[0027] Fig.10 Schematic diagrams of the structures of GRU and BiGRU of the present invention, wherein (a) is a schematic diagram of the structure of GRU, and (b) is a schematic diagram of the structure of BiGRU;

[0028] Fig.11 1 is a schematic diagram comparing the rule-based data processing results of the present invention with the unprocessed results, wherein (a) is a schematic diagram of the effect of the unprocessed data, and (b) is a schematic diagram of the effect of preliminary cleaning of the data;

[0029] Fig.12 It is a data processing result diagram based on the LOF algorithm of the present invention;

[0030] Fig.13 1 is a diagram of the early warning model result of the present invention, wherein (a) is a schematic diagram of the early warning result for the generator speed, and (b) is a schematic diagram of the early warning result for the temperature of the #3 variable pitch generator;

[0031] Fig.14 1 is a diagram of the early warning residual result of the present invention, wherein (a) is a schematic diagram of the early warning residual result 6.9 hours before the fault occurs, and (b) is a schematic diagram of the early warning residual result 3.6 hours before the fault occurs.

[0032] In conjunction with the accompanying drawings, the reference numerals in the embodiments of the present invention are as follows:

[0033] 102 - acquisition unit; 104 - control unit. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] Considering that the fault diagnosis methods for wind turbines in the relevant solutions are mainly concentrated in vibration signal analysis, data-driven fault diagnosis methods, etc., there is a lack of universality, and the accuracy of fault diagnosis of wind turbines under complex weather conditions cannot be guaranteed. In addition, in the relevant solutions, some innovative model architectures (such as Transformer-based models) are computationally complex, require high hardware resources, and are time-consuming to train. The methods that combine multiple technologies (such as Cheng Yuanjian's research) have complex model structures and algorithms due to the integration of multiple elements, and require high technical levels of operators, which increases the difficulty of practical application.

[0036] Therefore, the solution of the present invention proposes a fault warning method for a wind turbine, specifically a fault warning method for a wind turbine pitch system based on CycleNet and a multi-head attention mechanism. The use of the multi-head attention mechanism makes the model more efficient in learning feature states, and the use of a cyclic network can make feature learning more comprehensive. Under complex weather conditions, fault warning of the wind turbine pitch system can be achieved, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0037] According to an embodiment of the present invention, a fault warning method for a wind turbine is provided. Figure 1 The flow chart of an embodiment of the method of the present invention is shown in FIG. The fault warning method of the wind turbine generator set may include: steps S110 to S170.

[0038] In step S110 , historical data of the wind turbine generator set is collected from the SCADA system of the wind turbine generator set.

[0039] In step S120, historical data of the wind turbine generator system is processed to obtain samples, and the samples are divided into a training set and a test set.

[0040] At step S130, an improved BiGRU model is constructed based on CycleNet and the multi-head attention mechanism.

[0041] In step S140, the improved BiGRU model is trained using the training set to obtain an operation prediction model of the wind turbine generator set.

[0042] In step S150, after obtaining the operation prediction model of the wind turbine set, a preset sliding window is added to the operation prediction model of the wind turbine set, and the test set is used to perform testing to obtain test results; and the operation prediction model of the wind turbine set is corrected according to the test results to obtain a fault prediction model of the wind turbine set.

[0043] In step S160, current data of the wind turbine generator system is obtained.

[0044] In step S170, a fault prediction model of the wind turbine generator set is used to predict whether the wind turbine generator set is faulty based on current data of the wind turbine generator set, so as to implement fault warning for the wind turbine generator set.

[0045] The solution of the present invention proposes a fault warning scheme for a wind turbine pitch system based on CycleNet and a multi-head attention mechanism. The multi-head attention mechanism is used to make the model more efficient in learning feature states, and the use of a cyclic network can enable more comprehensive feature learning. Under complex weather conditions, fault warning of the wind turbine pitch system can be achieved, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0046] In some implementations, the specific process of processing the historical data of the wind turbine generator system to obtain samples in step S120 is described in the following exemplary embodiments.

[0047] Combine the following Figure 2 The flowchart of an embodiment of the method of the present invention for processing the historical data of the wind turbine to obtain samples further illustrates the specific process of processing the historical data of the wind turbine to obtain samples in step S120, including: step S210 to step S230.

[0048] Step S210: cleaning the historical data of the wind turbine generator set to obtain first processed data of the wind turbine generator set.

[0049] Step S220 , using a Pearson correlation algorithm, extracting characteristic data of the variable pitch system of the wind turbine generator set from the first processed data of the wind turbine generator set to obtain second processed data.

[0050] Step S230: After normalizing the second processed data, select the operating data of the wind turbine group as a sample.

[0051] Figure 8 FIG. 1 is a schematic diagram of a fault warning process in a fault warning method for a wind turbine pitch system based on CycleNet and a multi-head attention mechanism of the present invention. Figure 8 As shown, the fault warning process in the solution of the present invention includes:

[0052] Step 1: Extract data from the supervisory control and data acquisition (SCADA) system, process the extracted data, and select the normal operation data of the wind turbine to divide it into the training set and test set of the model.

[0053] For example, we use the historical operation data of two wind turbines numbered D18 and D23 in a wind farm in Xinjiang, which contains a large amount of normal operation data and some abnormal data due to shutdown caused by faults. First, we remove discrete points (usually points on the wind speed power curve that deviate significantly from the normal data distribution) and abnormal points based on basic rules based on expert experience. The rules are as follows: (1) The wind speed is less than the cut-in wind speed but the active power and generator speed are greater than 0; (2) There is an abnormal situation in the ratio of generator speed to active power, etc.; (3) Clean up abnormal values, duplicate values, missing values, and 0 values.

[0054] Step 2: Use the Pearson correlation algorithm to select the relevant features of the variable pitch system, and then normalize the data to eliminate the impact of different dimensions between different features.

[0055] In step 2, according to formula (6), the Pearson correlation coefficient method is used to select input features of the cleaned data: first, the characteristic variables related to the pitch system are selected, and then the features with strong correlation with the pitch system are screened out by the Pearson correlation coefficient method, and finally, the features such as generator speed, 3# pitch motor temperature, 2# pitch motor temperature, 1# pitch motor temperature, 2# pitch inverter temperature, 3# pitch inverter temperature, 2# unit grid-side inverter temperature, 2# unit motor-side inverter temperature, 1# unit grid-side inverter temperature are screened out as input variables.

[0056]

[0057] Among them, r is the Pearson correlation coefficient, X i and Y i is the ith observation of two variables X and Y (the number of observations is n), and is the average of the two variables. The characteristic variables selected using the Pearson algorithm are used as input variables of the prediction model.

[0058] Fig. 9 It is a schematic diagram of the data processing flow in a wind turbine variable pitch system fault warning method based on CycleNet and multi-head attention mechanism of the present invention. Fig. 9In view of the problem that SCADA data is huge and complex, according to the wind speed power curve and the generator speed curve, we first remove discrete points and abnormal points based on basic rules based on expert experience, such as the situation where the wind speed is less than the cut-in wind speed but the active power and generator speed are greater than 0, and the ratio of generator speed to active power has abnormal conditions. Then, we clean the abnormal values, duplicate values, missing values, and 0 values. After many experiments, we select the LOF algorithm to clean the data again, and finally use the min-max algorithm to normalize the data.

[0059] Compared with related solutions, the solution of the present invention, by applying Pearson correlation coefficient analysis and LOF algorithm, we thoroughly clean the data and select features, ensuring the data quality and relevance of the input model, thereby improving the accuracy of the model.

[0060] In some implementations, the specific process of cleaning the historical data of the wind turbine generator set to obtain the first processed data of the wind turbine generator set in step S210 is described in the following exemplary embodiments.

[0061] Combine the following Figure 3 The flowchart of an embodiment of cleaning the historical data of the wind turbine generator set in the method of the present invention further illustrates the specific process of cleaning the historical data of the wind turbine generator set in step S210, including: step S310 to step S320.

[0062] Step S310 , cleaning abnormal data in historical data of the wind turbine generator system according to a preset abnormal data range, to obtain first cleaned data.

[0063] Step S320: Use the LOF algorithm to clean the discrete points and abnormal data in the first cleaned data again to obtain first processed data of the wind turbine generator set.

[0064] like Figure 8 As shown, the fault warning process in the scheme of the present invention also includes: in step 1, in the operation of processing the extracted data, firstly, the extracted data is cleaned for abnormal data using a rule-based method, wherein the rule-based method is as follows: 1. The wind speed is less than the cut-in wind speed but the active power and the generator speed are greater than 0; 2. There are abnormal conditions in the ratio of the generator speed to the active power, etc.; 3. Clean the abnormal values, duplicate values, missing values, and 0 values. Then use the local outlier factor (LOF) algorithm to clean the discrete points and abnormal data. Among them, the LOF algorithm is used to identify the abnormal values ​​by calculating the local density deviation of a given data point relative to its neighbors to clean the data again. After cleaning in step 1, the data samples are 194,354, and after cleaning by the LOF algorithm, the data samples are 154,511.

[0065] When the LOF algorithm is used to calculate the local density deviation of a given data point relative to its neighbors, the k neighbor distances are calculated: for each data point cleaned in step 1, its k nearest neighbor data points are found, and the distance between each data point and the k nearest neighbor data points is calculated according to formulas (1) and (2):

[0066] d k (P) = d(P,O) (1);

[0067]

[0068] Where d(P,O) is the Euclidean distance between point P and point O, d k (P) is the k-distance of P, where k is an important parameter that determines the calculation range of the local density; d Euclid (X i ,X j ) is represented by X i and X j The Euclidean distance between .

[0069] When using the LOF algorithm to calculate the local density deviation of a given data point relative to its neighbors, the reachable distance is also calculated: for each data point, the reachable distance to other points in the k-distance domain is calculated according to formula (3):

[0070] N k (P)={d(P,O')≤d k (P)} (3).

[0071] Where d(P,O) is the Euclidean distance between point P and point O, d k (P) is the K-neighbor distance of data point P, N k (P) represents the K nearest neighbor set of sample point P.

[0072] When using the LOF algorithm to calculate the local density deviation of a given data point relative to its neighbors, the local reachable density is also calculated: for each data point, the local reachable density within its k-distance domain is calculated according to formula (4):

[0073]

[0074] Among them, lrd k (P) is the local reachability density of the k nearest neighbors of point P. N k (P) is the K nearest neighbor set of data point P, |N k (P)| is the number of neighbors of P.

[0075] When using the LOF algorithm to calculate the local density deviation of a given data point relative to its neighbors, the LOF value is also calculated: For each data point, its LOF value is calculated according to formula (5), which is obtained by comparing the local reachable density of the point with the average level of the local reachable density of its k neighbors:

[0076]

[0077] Where d(P,O) is the Euclidean distance between point P and point O; lrd k and lrd(O) is the local reachability density of point P and its neighbor point O; LOF k The (P) value is calculated by comparing the local reachability density of point P with the average level of the local reachability densities of its k neighbors.

[0078] Formula (1) is the basis for calculating all other metrics because it defines the range of the local neighborhood; Formula (2) is based on the k-nearest neighbor distance, which takes into account the actual distance from point P to its nearest neighbor point O and the local density of point P; Formula (3) is based on the reachable distance calculation, which reflects the density in the local area of ​​point P; Formula (4) compares the local reachable density of point O with the local reachable density of point P to measure the density difference of O relative to P; Formula (5) combines the relative reachable density of all neighboring points of point P to provide an indicator to measure whether point P is an outlier. The higher the LOF value, the more likely point P is an outlier. Through these formulas, the LOF algorithm can identify data points whose density in their local neighborhood is significantly lower than that of neighboring points, thereby marking them as outliers.

[0079] Fig.11 It is a schematic diagram comparing the rule-based data processing results of the present invention with the unprocessed results, wherein (a) is a schematic diagram of the effect of the unprocessed data, and (b) is a schematic diagram of the effect of preliminary cleaning of the data. Fig.11 This is a comparison chart of the rule-based data processing results and the unprocessed data of the present invention. The data is processed according to the wind speed power curve, and the unprocessed data is shown in (a). First, a series of rules are formulated using expert experience to perform preliminary cleaning of the data as shown in (b).

[0080] Fig.12 This is a diagram of data processing results based on the LOF algorithm of the present invention. Fig.12 This is a diagram of data processing results based on the LOF algorithm of the present invention. The pre-processed data is cleaned again using the LOF algorithm to remove noise and some discrete points.

[0081] In some embodiments, the specific process of constructing an improved BiGRU model based on CycleNet and the multi-head attention mechanism in step S130 is shown in the following exemplary description.

[0082] Combine the following Figure 4 The flowchart of an embodiment of the method of the present invention for constructing an improved BiGRU model based on CycleNet and multi-head attention mechanism is shown, which further illustrates the specific process of constructing the improved BiGRU model based on CycleNet and multi-head attention mechanism in step S120, including: step S410 to step S450.

[0083] Step S410, based on BiGRU, obtain the forward GRU part and the reverse GRU part.

[0084] Step S420, using the CycleNet network, merge the forward GRU part and the reverse GRU part to obtain a merged model.

[0085] Step S430, after adding a dropout layer to the merged model, the features in the merged model are mapped to the output space through a fully connected layer.

[0086] Step S440: In the output space, based on the multi-head attention mechanism, the input of the output space is divided into more than two heads, and an attention matrix is ​​calculated for each head.

[0087] Step S450, the attention matrices of more than two heads are connected in series to form a new matrix as the BiGRU model.

[0088] Among them, GRU is the basic unit of BiGRU. The Cyclenet model is a recurrent network that cyclically learns the features learned by the two GRU units in BiGRU. Its function is to avoid losing some key information and better learn the information in the features. The model finally obtained by the scheme of the present invention is based on the BiGRU model, and the Cyclenet recurrent network is used as an improvement, and then a model with a multi-head attention mechanism is combined.

[0089] like Figure 8 As shown, the fault warning process in the solution of the present invention also includes:

[0090] Step 3: Build a model that includes a bidirectional gated recurrent unit (BiGRU) and a multi-head attention mechanism. The forward GRU part of the model starts from the first time step of the sequence and gradually calculates the state of each time step; the reverse GRU part of the model starts from the last time step of the sequence and gradually calculates the state of each time step in reverse; CycleNet is used to capture sequence features over a long period of time; then the outputs of the forward and reverse GRUs are merged so that the state of each time step contains both forward and backward information. A dropout layer is added to the model to reduce the risk of overfitting; and the merged features are then mapped to the output space through a fully connected layer. In the multi-head attention mechanism, the input is divided into multiple heads, each head calculates its own attention matrix; the attention matrix results of all heads are connected in series to form a new matrix. Among them, the features merged by the fully connected layer are the features learned by the forward and backward GRU units and the features learned using the Cyclenet recurrent network. The main purpose of using a multi-head attention mechanism to split into multiple heads is to save computing time due to the large amount of data.

[0091] In step 3, a model including a bidirectional gated recurrent unit (BiGRU) and a multi-head attention mechanism is constructed. By merging the outputs of the forward and reverse GRUs, the state of each time step contains both forward and backward information. The four main calculation processes of the GRU include the calculation of the update gate, the calculation of the reset gate, the calculation of the candidate hidden state, and the calculation of the final hidden state.

[0092] Among them, the update gate determines to what extent the previous hidden state information is retained when calculating the hidden state at the current moment, as shown in formula (7):

[0093] z t =σ(W z [h t-1 ,x t ]+b z ) (7).

[0094] Among them, z t is the output of the update gate at time step t; σ is the sigmoid activation function, and the output range of σ is [0,1], which is used to determine the degree of retaining old information; W z is the weight matrix of the update gate; h t-1 is the hidden state at the previous moment; x t is the input at the current moment; b z is the bias term of the update gate.

[0095] The reset gate determines to what extent the previous hidden state information is ignored when calculating the candidate hidden state at the current moment, as shown in formula (8):

[0096] rt =σ(W r [h t-1 ,x t ]+b r ) (8).

[0097] Among them, r t is the output of the reset gate at time step t, σ is the sigmoid activation function, and the output range of σ is [0,1], which is used to determine the degree of retaining old information; W r is the weight matrix of the update gate; h t-1 is the hidden state at the previous moment; x t is the input at the current moment; b r is the bias term of the update gate.

[0098] The candidate hidden state is calculated based on the current input and part of the previous hidden state information, which takes into account the influence of the reset gate, as shown in formula (9):

[0099]

[0100] Among them, h t ′ is the candidate hidden state at time step t; tanh is the hyperbolic tangent activation function, whose output range is [-1,1]; W is the weight matrix of the hidden state; h t-1 is the hidden state at the previous moment; x t is the input at the current moment; b is the bias term of the hidden state.

[0101] The final hidden state is calculated by combining the update gate and the candidate hidden state, which determines the degree of fusion of new information and old information, as shown in formula (10):

[0102]

[0103] Among them, h t is the hidden state at time t; z t is the output of the update gate at time step t; h t-1 is the hidden state at the previous moment; h t ′ is a candidate hidden state at time step t.

[0104] BiGRU consists of two GRU units, one for processing the forward sequence and the other for processing the reverse sequence. 1 ,x 2 ,...,x T ], calculate the forward GRU hidden state Compute the inverse GRU hidden state Finally, the forward and reverse hidden states are merged, and the merged hidden state is expressed as formula (11):

[0105]

[0106] Fig.10 Schematic diagrams of the structures of GRU and BiGRU of the present invention, wherein (a) is a schematic diagram of the structure of GRU, and (b) is a schematic diagram of the structure of BiGRU. Fig.10 It is a structural diagram of GRU and BiGRU of the present invention, and BiGRU is used to perform feature learning on data processed by bidirectional GRU units. Fig.10 (a) is the GRU structure diagram, and (b) is the BiGRU structure diagram.

[0107] In step 2, in order to make the data comparable between different features in view of the influence between different dimensions, the data is normalized by min-max, as shown in formula (12):

[0108]

[0109] Where x is the normalized data, x i is the original data, x min is the smallest data in the original data, x max It is the largest data in the original data. In the solution of the present invention, after extracting features using the Pearson algorithm, the selected feature data is normalized, that is, step 2.

[0110] Compared with related solutions, the solution of the present invention introduces CycleNet, which is a variant of the recurrent neural network (RNN) architecture. It captures sequence features over a long period of time through a cyclic structure, enhances the model's ability to recognize complex patterns, and improves the time series processing capability.

[0111] In some embodiments, after adding a preset sliding window to the operation prediction model of the wind turbine in step S140, the specific process of using the test set to perform testing to obtain the test results is described in the following exemplary embodiments.

[0112] Combine the following Figure 5 The flowchart of an embodiment of using the test set to perform testing to obtain test results in the method of the present invention further illustrates the specific process of using the test set to perform testing to obtain test results in step S140, including: steps S510 to S520.

[0113] Step S510, after the operation prediction model of the wind turbine is constructed, a sliding window statistical method is adopted, a preset sliding window is added, and fault warning is performed on the data of the test set to obtain a warning result.

[0114] Step S520, according to the warning result, the upper and lower bounds of the residual threshold of the operation prediction model of the wind turbine are set as an indicator for confirming whether to issue a warning for the output result of the operation prediction model of the wind turbine, and the indicator is used as the test result.

[0115] like Figure 8 As shown, the fault warning process in the solution of the present invention also includes:

[0116] Step 4: Use mean square error MSE, squared absolute error MAE, and determination coefficient R as the evaluation indicators of the model 2 That is to say, the performance evaluation index of the fault warning model of the variable pitch system is the mean square error MSE, the square absolute error MAE, and the determination coefficient R 2 . The mean square error (MSE) is the average of the squares of the differences between the predicted value and the true value. It measures the mean of the sum of the squares of the deviations between the predicted value and the true value, as shown in formula (13). The mean absolute error (MAE) is the average of the absolute values ​​of the differences between the predicted value and the true value. It directly measures the average deviation between the predicted value and the true value, as shown in formula (14). The coefficient of determination (R) 2 It reflects the degree to which the independent variable explains the dependent variable. Its value range is [0,1]. The closer it is to 1, the better the regression model fits the data. For details, see formula (15).

[0117]

[0118] Where n is the number of samples, y i is the ith actual value, is the ith predicted value, is the average of all actual values, is the square of the ith prediction error, is the absolute value of the ith prediction error, is the sum of squares of the errors between the predicted value and the actual value. is the sum of squares of the errors between the actual value and the mean value.

[0119] Step 5: After building the normal operation prediction model, add the sliding window method to perform fault warning on the data of the validation set, and set the upper and lower bounds of the residual threshold as indicators for warning.

[0120] In step 5, the sliding window statistical method is used to determine the warning value. Within a certain period of time, the residual sequence of the fault warning model is formula (16):

[0121] ε=[ε 1 , ε 2 , ε n …] (16).

[0122] In formula (16), ε is the difference sequence between the actual value and the predicted value of the residual, ε 1 ,ε 2 ,ε n ···is the difference between the 1st, 2nd, and nth actual values ​​and the predicted values.

[0123] Take a sliding window with a width of N for the residual sequence, where the value of N is determined based on field experience. The standard deviation of N consecutive residuals in the window is η, where is the mean of the residuals, as shown in formula (17):

[0124]

[0125] The threshold is determined based on the statistical characteristics of the residual data in the sliding window. This method takes into account the local volatility of the data. By dynamically setting the threshold according to the data distribution in the window, it can better adapt to the possible changes in the data in different local areas. The threshold α is expressed as formula (18):

[0126]

[0127] Among them, K is the calculation coefficient and is generally greater than 1, determined by field experience, and η is the standard deviation.

[0128] Compared with related solutions, the solution of the present invention uses a sliding window method to monitor fault characteristics in real time, so that the model can focus on local fragments of data and capture local sequence patterns. As new data continues to enter, the model can update the data set for early warning judgment in a timely manner, significantly improving the real-time and accuracy of early warning.

[0129] In the scheme of the present invention, the data of the data acquisition and monitoring control (SCADA) system is first extracted, the data is cleaned, and the relevant features of the variable pitch system are selected using the Pearson correlation algorithm. Then the data is normalized to eliminate the influence of different dimensions between different features, and the normal operation data of the wind turbine is selected to be divided into the training set and test set of the model; a model including a bidirectional gated recurrent unit (BiGRU) and a multi-head attention mechanism is constructed, and CycleNet is added to capture the sequence characteristics in a long period of time, and then the outputs of the forward and reverse GRUs are merged so that the state of each time step contains forward and backward information. A dropout layer is added to the model to reduce the risk of overfitting. Among them, the selection of normal operation data is to first establish a normal operation prediction model: the model learns the normal operation data features and predicts the next operation data value based on the normal operation data, and takes the operation data of the wind turbine with faults as input. If the difference between the actual value and the predicted value deviates from the threshold range, the sliding window method will be used to implement early warning. The evaluation indicators of the model use mean square error MSE, square absolute error MAE, and determination coefficient R 2 After building the normal operation prediction model, the sliding window method is added to perform fault warning on the data of the validation set. By setting the upper and lower bounds of the residual threshold as indicators of whether to issue a warning, fault warning of the wind turbine variable pitch system can be achieved under complex weather conditions, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0130] In some implementations, the fault warning method for a wind turbine set described in the solution of the present invention further includes: a process of optimizing an operation prediction model of the wind turbine set.

[0131] Combine the following Figure 6 The flowchart of an embodiment of optimizing the operation prediction model of the wind turbine generator set in the method of the present invention further illustrates the specific process of optimizing the operation prediction model of the wind turbine generator set, including: step S610 to step S630.

[0132] Step S610, after obtaining the operation prediction model of the wind turbine, use the crown porcupine optimization algorithm to optimize the hyperparameters of the operation prediction model of the wind turbine to obtain a CPO-BiGRU model constructed based on CycleNet and a multi-head attention mechanism, which is recorded as the optimized CPO-BiGRU model.

[0133] Step S620: training the optimal CPO-BiGRU model using the training set to obtain an optimal operation prediction model for the wind turbine.

[0134] Step S630, after obtaining the optimal operation prediction model of the wind turbine group, after adding a preset sliding window to the optimal operation prediction model of the wind turbine group, use the test set to test to obtain a test result; and according to the test result, correct the optimal operation prediction model of the wind turbine group to obtain the optimized fault prediction model of the wind turbine group; then, obtain the current data of the wind turbine group; use the fault prediction model of the wind turbine group to predict whether the wind turbine group is faulty based on the current data of the wind turbine group, so as to realize fault warning of the wind turbine group.

[0135] like Figure 8 As shown, the fault warning process in the solution of the present invention also includes: step 6, using the crown porcupine optimization algorithm (CPO) to optimize the hyperparameters, specifically: calculating the fitness of each crown porcupine and checking whether it meets the preset threshold, calculating the fitness of each candidate solution and checking whether it meets the preset threshold; updating the position of the crown porcupine, and further optimizing the search space to find a better solution in the search space.

[0136] In step 6, hyperparameter optimization is to optimize the hyperparameters, specifically: using the crown porcupine optimization algorithm (CPO) to optimize the hyperparameters of the model to improve the prediction accuracy of the model.

[0137] Accelerate the convergence speed while maintaining the diversity of the group. The algorithm accelerates the convergence speed by letting some crested porcupines (CP) leave the group during the optimization process, and then rejoin the group to improve diversity and avoid falling into the local minimum. The mathematical model of cyclic group size reduction is as shown in formula (19):

[0138]

[0139] Among them, N represents the number of initial search agents, that is, the population size, N′ represents the number of individuals, T is the variable that determines the number of loops, t is the current function evaluation, and T max is the maximum number of function evaluations, N min is the minimum number of individuals in the newly generated population.

[0140] The crested porcupine optimization algorithm simulates the foraging behavior of the crested porcupine and can perform global search in a wide search space to avoid falling into local optimal solutions. This global search capability helps to find better hyperparameter combinations and improve the performance of the model. It can also perform efficient optimization for wind turbine variable pitch systems with multiple feature inputs. Compared with other complex optimization algorithms, the crested porcupine optimization algorithm is relatively simple to implement and its parameters are relatively easy to adjust. In the initialization stage, the optimization problem consists of multiple dimensions (d), and each dimension needs to be minimized or maximized through optimization techniques. The CPO algorithm optimizes these problems by randomly initializing their dimensions in the search space of each problem. The CPO algorithm is as shown in formula (20):

[0141] X i =L+r×(UL) i=1,2,...,N′ (20).

[0142] Where N′ represents the number of individuals, X i is the i-th candidate solution in the search space, L and U are the lower and upper bounds of the search range, respectively, and r is a vector randomly initialized between [0,1].

[0143] Cyclic population reduction technique is used to maintain the diversity of the population while accelerating convergence. This strategy simulates that not all CPs activate the defense mechanism, but only those CPs that are threatened will activate the defense mechanism. During the optimization process, this method takes some CPs out of the population to speed up convergence and puts them back into the population to improve diversity and avoid falling into local minima. The specific formula is (21):

[0144]

[0145] Among them, T is the variable that determines the number of loops, t is the current function evaluation, and T max is the maximum number of function evaluations, N min is the minimum number of individuals in the newly generated population.

[0146] According to the defensive behavior of CP, when the predator is far away, CP has two defense strategies, namely visual strategy and sound strategy. These strategies refer to investigating different areas and devoting themselves to global exploration, as shown in formula (22):

[0147]

[0148] Among them, x CP is the optimal CP position, y t is the vector generated between the current CP and the random CP, is the position vector of the i-th search agent in the t+1th iteration, is the position vector of the i-th search agent in the t-th iteration, τ 1 , τ 2 is a random number used to control the update behavior of the search agent. The following symbols numbered with τ are all random numbers used to control the update behavior of the search agent:

[0149]

[0150] in, is a randomly selected individual position in the population, is the position vector of the i-th search agent in the t-th iteration, Used to represent an intermediate position reference between the current individual and the random individual during iteration.

[0151] CP uses sound to make noise and threaten predators. When the predator approaches, the sound of CP becomes louder, as shown in formula (23):

[0152]

[0153] in, and Respectively represent the positions of r1 and r2 individuals in the population, Used to control the balance between exploration and exploitation during search agent updates.

[0154] According to the defensive behavior of CP, when a predator approaches, CP has two defense strategies, namely, the odor attack strategy and the physical attack strategy. These strategies refer to the development of potential areas and the focus on (local) development search. In the odor attack phase, CP will secrete a foul odor that spreads in the surrounding area to prevent predators from approaching it, as shown in formula (24):

[0155]

[0156] in, is a parameter that controls the search direction, γ t As a defense factor, is the odor diffusion factor, is the objective function value of the ith individual in the tth iteration, is the objective function value of all search agents in the tth iteration, t max is the maximum number of function evaluations, is a random vector between [0,1], and rand is a random number between [0,1];

[0157] In the physical attack phase, when the predator is very close to it, the CP will attack it with short and thick feathers, as shown in formula (25):

[0158]

[0159] Where α is the convergence speed factor, It is a fitness-based factor that affects the updating of the search agent.

[0160] Compared with related solutions, the solution of the present invention adopts the crested porcupine optimization algorithm (CPO) to optimize the hyperparameters of the model. The CPO that simulates the foraging behavior of crested porcupines can perform a global search in a wide search space to avoid falling into a local optimal solution, thereby finding a better hyperparameter combination and further improving the model prediction performance.

[0161] Fig.13 It is a diagram of the early warning model result of the present invention, wherein (a) is a schematic diagram of the early warning result for the generator speed, and (b) is a schematic diagram of the early warning result for the temperature of the #3 variable pitch generator. Fig.13 This is the result diagram of the early warning model of the present invention. For the generator speed as shown in Figure (a) and the temperature of the #3 variable pitch generator as shown in Figure (b), the curves of the predicted value and the actual value are approximately consistent. Therefore, the result diagram can illustrate that the constructed normal operation prediction model has a good prediction ability.

[0162] Fig.14 1 is a diagram of the early warning residual result of the present invention, wherein (a) is a schematic diagram of the early warning residual result 6.9 hours before the fault occurs, and (b) is a schematic diagram of the early warning residual result 3.6 hours before the fault occurs. Fig.14 This is the early warning residual result diagram of the present invention. It warns of the two 3# variable pitch inverter OK signal loss faults and 1#U3_ hardware overcurrent faults. Warnings are given 6.9 hours (a) and 3.6 hours (b) before the faults occur. Warnings are given again multiple times before the faults occur, leaving ample time for maintenance and inspection of the equipment. The superiority of the model is verified. The red lines in the figure are the upper and lower limits of the thresholds, and the black lines are the residual values.

[0163] Compared with the related solutions, the solution of the present invention not only provides fault warning, but also can comprehensively evaluate the warning results. By comparing with the actual fault occurrence, the model parameters are continuously adjusted and optimized to improve the accuracy and reliability of the warning.

[0164] By adopting the technical solution of this embodiment, historical data of the wind turbine is extracted, and after cleaning, feature selection, and normalization processing of the historical data, the wind turbine operation data in the wind turbine is selected as a sample, and the sample is divided into a training set and a test set; an improved BiGRU model including a bidirectional gated recurrent unit and a multi-head attention mechanism is constructed, and after training with the training set, the model is tested and corrected with the test set to obtain a fault warning model; the current data of the wind turbine is obtained to use the fault warning model to predict whether the wind turbine is faulty based on the current data; thus, the fault warning model is obtained by learning based on CycleNet and the multi-head attention mechanism, and the fault warning of the wind turbine variable pitch system can be realized under complex weather conditions, which is beneficial to improving the operating efficiency and reliability of wind power equipment.

[0165] According to an embodiment of the present invention, a fault warning device for a wind turbine generator set corresponding to the fault warning method for the wind turbine generator set is also provided. Figure 7 The structural schematic diagram of an embodiment of the device of the present invention is shown. The fault warning device of the wind turbine generator set may include: an acquisition unit 102 and a control unit 104 .

[0166] The acquisition unit 102 is configured to collect the historical data of the wind turbine from the SCADA system of the wind turbine. The specific functions and processing of the acquisition unit 102 refer to step S110.

[0167] The control unit 104 is configured to process the historical data of the wind turbine to obtain samples, and divide the samples into a training set and a test set. The specific functions and processing of the control unit 104 refer to step S120.

[0168] The control unit 104 is further configured to construct an improved BiGRU model based on CycleNet and the multi-head attention mechanism. The specific functions and processing of the control unit 104 are also shown in step S130.

[0169] The control unit 104 is further configured to train the improved BiGRU model using the training set to obtain the operation prediction model of the wind turbine. The specific functions and processing of the control unit 104 are also shown in step S140.

[0170] The control unit 104 is further configured to, after obtaining the operation prediction model of the wind turbine generator set, add a preset sliding window to the operation prediction model of the wind turbine generator set, perform a test using the test set to obtain a test result; and modify the operation prediction model of the wind turbine generator set according to the test result to obtain a fault prediction model of the wind turbine generator set. The specific functions and processing of the control unit 104 are also shown in step S150.

[0171] The acquisition unit is further configured to acquire current data of the wind turbine. The specific functions and processing of the acquisition unit 102 are also shown in step S160.

[0172] The control unit is further configured to use the fault prediction model of the wind turbine to predict whether the wind turbine is faulty based on the current data of the wind turbine, so as to implement fault warning for the wind turbine. The specific functions and processing of the control unit 104 are also shown in step S170.

[0173] In the solution of the present invention, a multi-head attention mechanism is used to make the model more efficient in learning feature states, and a recurrent network is used to learn features more fully, which can realize fault warning of the wind turbine pitch system under complex weather conditions, and is conducive to improving the operating efficiency and reliability of wind power equipment.

[0174] Since the processing and functions implemented by the device of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0175] According to an embodiment of the present invention, a wind turbine generator set corresponding to the fault warning device of the wind turbine generator set is also provided. The wind turbine generator set may include: the fault warning device of the wind turbine generator set described above.

[0176] Since the processing and functions implemented by the wind turbine generator set of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned device, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0177] According to an embodiment of the present invention, a storage medium corresponding to a fault warning method for a wind turbine is also provided, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the above-mentioned fault warning method for the wind turbine.

[0178] Since the processing and functions implemented by the storage medium of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0179] According to an embodiment of the present invention, a computer program product corresponding to the fault warning method for a wind turbine is also provided, comprising a computer program, which implements the steps of the above-mentioned fault warning method for a wind turbine when executed by a processor.

[0180] Since the processing and functions implemented by the product of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned terminal, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0181] In summary, it is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned advantageous methods can be freely combined and superimposed.

[0182] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A fault warning method for a wind turbine generator set, characterized in that: include: Collecting historical data of the wind turbine generator set from the SCADA system of the wind turbine generator set; Processing the historical data of the wind turbine generator set to obtain samples, and dividing the samples into a training set and a test set; Based on CycleNet and multi-head attention mechanism, an improved BiGRU model is constructed; Using the training set to train the improved BiGRU model, to obtain an operation prediction model of the wind turbine generator set; After adding a preset sliding window to the operation prediction model of the wind turbine generator set, the test set is used to perform a test to obtain a test result; and modifying the operation prediction model of the wind turbine generator set according to the test results to obtain a fault prediction model of the wind turbine generator set; Acquiring current data of the wind turbine generator set; The fault prediction model of the wind turbine generator set is used to predict whether the wind turbine generator set is faulty based on current data of the wind turbine generator set, so as to achieve fault warning for the wind turbine generator set.

2. The fault warning method for a wind turbine generator set according to claim 1, characterized in that: The historical data of the wind turbine generator set is processed to obtain samples, including: Cleaning the historical data of the wind turbine generator set to obtain first processed data of the wind turbine generator set; Using a Pearson correlation algorithm, extracting characteristic data of a variable pitch system of the wind turbine generator set from the first processed data of the wind turbine generator set to obtain second processed data; After normalizing the second processed data, the operating data of the fan unit is selected as a sample.

3. The fault warning method for a wind turbine generator set according to claim 2, characterized in that: Cleaning the historical data of the wind turbine generator set to obtain first processed data of the wind turbine generator set includes: According to a preset abnormal data range, the abnormal data in the historical data of the wind turbine generator set is cleaned to obtain first cleaned data; The discrete points and abnormal data in the first cleaned data are cleaned again using the LOF algorithm to obtain first processed data of the wind turbine generator set.

4. The fault warning method for a wind turbine according to any one of claims 1 to 3, characterized in that: Based on CycleNet and multi-head attention mechanism, an improved BiGRU model is constructed, including: Based on BiGRU, we get the forward GRU part and the reverse GRU part; Using a CycleNet network, the forward GRU part and the reverse GRU part are merged to obtain a merged model; After adding a dropout layer to the merged model, the features in the merged model are mapped to the output space through a fully connected layer; In the output space, based on a multi-head attention mechanism, the input of the output space is divided into two or more heads, and an attention matrix is ​​calculated for each head; The attention matrices of more than two heads are concatenated to form a new matrix as the CycleNet model.

5. The fault warning method for a wind turbine according to any one of claims 1 to 4, characterized in that: After adding a preset sliding window to the operation prediction model of the wind turbine generator set, the test set is used to perform a test to obtain a test result, including: After the operation prediction model of the wind turbine is constructed, a sliding window statistical method is used, a preset sliding window is added, and a fault warning is performed on the data of the test set to obtain a warning result; According to the warning result, the upper and lower bounds of the residual threshold of the operation prediction model of the wind turbine group are set as an indicator for confirming whether to issue a warning for the output result of the operation prediction model of the wind turbine group, and the indicator is used as the test result.

6. The fault warning method for a wind turbine according to any one of claims 1 to 5, characterized in that: Also includes: After obtaining the operation prediction model of the wind turbine generator set, the crown porcupine optimization algorithm is used to optimize the hyperparameters of the operation prediction model of the wind turbine generator set to obtain a CPO-BiGRU model based on CycleNet and a multi-head attention mechanism, which is recorded as an optimized CPO-BiGRU model; Using the training set to train the optimal CPO-BiGRU model, to obtain an optimal operation prediction model for the wind turbine generator set; After adding a preset sliding window to the optimal operation prediction model of the wind turbine generator set, the test set is used to perform a test to obtain a test result; The optimized operation prediction model of the wind turbine generator set is corrected according to the test results to obtain an optimized fault prediction model of the wind turbine generator set.

7. A fault warning device for a wind turbine generator set, characterized in that: include: An acquisition unit is configured to collect historical data of the wind turbine generator set from a SCADA system of the wind turbine generator set; A control unit, configured to process the historical data of the wind turbine generator set to obtain samples, and divide the samples into a training set and a test set; The control unit is further configured to construct a BiGRU model based on CycleNet and a multi-head attention mechanism; The control unit is further configured to train the improved BiGRU model using the training set to obtain an operation prediction model of the wind turbine generator set; The control unit is further configured to, after adding a preset sliding window to the operation prediction model of the wind turbine generator set, perform a test using the test set to obtain a test result; and modifying the operation prediction model of the wind turbine generator set according to the test results to obtain a fault prediction model of the wind turbine generator set; The acquisition unit is further configured to acquire current data of the wind turbine generator set; The control unit is further configured to use a fault prediction model of the wind turbine set to predict whether the wind turbine set is faulty based on current data of the wind turbine set, so as to implement fault warning for the wind turbine set.

8. A wind turbine generator set, characterized in that: include: A fault warning device for a wind turbine generator set as claimed in claim 7.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the fault warning method for a wind turbine set according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the fault warning method for a wind turbine set described in any one of claims 1 to 6 are implemented.

Images