Wind power plant unit rotating speed abnormity early warning method based on fusion model

Abstract

The invention discloses a wind power plant unit rotating speed abnormity early warning method based on a fusion model, and the method comprises the following steps: 1, obtaining the operation data of a certain group of wind power units, and carrying out the data preprocessing, normalization, correlation analysis and feature screening of the operation data, and obtaining an input data set; 2, constructing a fusion model, inputting the input data set into the fusion model for training, and optimizing the prediction effect of the fusion model by optimizing hyper-parameters to obtain an optimal fusion model; and step 3, through the optimal fusion model, predicting the average generator rotating speed of the wind turbine generator set to obtain a predicted value, calculating an error between the predicted value and a true value, judging the error based on a preset threshold value, and triggering early warning when abnormal signal accumulation exceeds a set condition to complete rotating speed abnormity early warning of the wind turbine generator set. The galloping early warning and monitoring accuracy can be improved, and the probability of early galloping accidents of the wind turbine generator and the maintenance cost are effectively reduced.

Description

Technical Field

[0001] This invention pertains to the field of fault diagnosis technology for offshore wind turbine generator systems, specifically relating to a method for early warning of abnormal wind turbine rotation speed based on a fusion model. Background Technology

[0002] Wind power, with its abundant resources and near-zero carbon emissions, plays a crucial role in the global energy transition, and its installed capacity and actual application scale have maintained rapid expansion in recent years. Runaway accidents are a typical high-risk safety incident threatening the safe operation of wind turbines. From the perspective of the mechanism, the causes and development logic of runaway accidents are relatively clear: on the one hand, after long-term operation, core components such as bearings and gearboxes may experience normal aging and wear; on the other hand, encountering extreme wind speeds or other harsh operating conditions may also impact the key systems of the unit. Both of these situations may lead to the failure of the unit's braking system or pitch system—and as the core component for controlling the unit's speed, once the pitch system fails, it becomes impossible to control the speed by adjusting the blade angle, ultimately causing the rotor speed to far exceed the rated value, putting the unit in a completely uncontrollable state. Such accidents are highly destructive: on the one hand, after an accident, maintenance personnel have difficulty quickly reaching the site to carry out emergency rescue, which can easily cause direct damage to core components such as generators and gearboxes, leading to long-term unit shutdowns; on the other hand, in severe cases, it can also cause catastrophic consequences such as nacelle fires and turbine collapse, not only causing huge property losses but also potentially posing a direct threat to the lives of maintenance personnel. From the current state of technological application, there are still some shortcomings in the overspeed warning and monitoring schemes for wind turbines: firstly, the accuracy of the warnings needs to be improved; existing models often fail to effectively capture the weak signals in the early stages of abnormal speed, easily leading to false alarms or missed alarms, and failing to provide a reliable early intervention window for maintenance work; secondly, the monitoring cost is relatively high, with some technical solutions relying on complex sensor arrays or additional dedicated monitoring equipment, the deployment, calibration, and subsequent maintenance costs of which are high, making it difficult to achieve widespread application in large-scale offshore wind farms. Therefore, developing a high-accuracy and cost-effective wind turbine speed anomaly warning method has become one of the important requirements for ensuring the safe and efficient operation of offshore wind power. Summary of the Invention

[0003] The purpose of this invention is to provide a method for early warning of abnormal wind turbine rotation speed based on a fusion model. This invention can improve the accuracy of overspeed warning and monitoring, and reduce the potential risks and maintenance costs of wind turbine overspeed monitoring.

[0004] The technical solution of this invention: a wind farm turbine speed anomaly early warning method based on a fusion model, comprising the following steps: Step 1: Collect multi-dimensional operation data of the SCADA system of a certain group of wind turbine units, and perform data preprocessing, normalization, correlation analysis and feature filtering on the operation data to obtain the input dataset; Step 2: Construct a fusion prediction model composed of a temporal convolutional network, an efficient channel attention mechanism, and a bidirectional long short-term memory network. Input the input dataset into the fusion model for training. Optimize the hyperparameters to achieve the best prediction effect and obtain the optimal fusion model. Step 3: The average generator speed of the target wind turbine is predicted using the optimal fusion model to obtain the speed prediction result. The deviation between the speed prediction result and the actual speed result is calculated. The deviation is judged based on a preset threshold. When the abnormal signal accumulates to exceed the set condition, an early warning is triggered to complete the wind turbine speed abnormality early warning.

[0005] In the aforementioned wind farm turbine speed anomaly early warning method based on fusion model, in step one, the characteristic parameters in the multi-dimensional operation data of the SCADA system include average generator speed, average wind speed, average grid active power, and average nacelle temperature.

[0006] In the aforementioned wind farm turbine speed anomaly early warning method based on fusion model, the data preprocessing includes removing feature parameters in the operating data whose proportion of zero and missing values ​​exceeds a threshold, and deleting data records in the operating data where the wind turbine is in an abnormal operating state. The normalization process applies a minimum-maximum normalization method to the remaining running data, transforming the feature dataset of the running data through a linear transformation. Any original feature parameters in The mapping is over the interval [0,1], and the specific formula is as follows: ; in, This represents the feature parameters after normalization. Representing the feature dataset The minimum value of all characteristic parameters in. Representing the feature dataset The maximum value of all feature parameters in the data; The correlation analysis uses the Pearson correlation coefficient method to calculate the characteristic parameters. By analyzing the correlation coefficients between the parameters, we can select the feature parameters that are highly correlated with the average generator speed, and obtain the input dataset. .

[0007] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, in step two, the fusion model is represented as: ; ; ; ; in, Indicates the input dataset. Represents a temporal convolutional network. Temporal features extracted by a temporal convolutional network. This indicates an efficient channel attention mechanism. Features weighted for efficient channel attention mechanisms It is a bidirectional long short-term memory network. The hidden state features output by BiLSTM. It is a fully connected layer. These are the predicted values ​​from the fusion model.

[0008] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, the temporal convolutional network includes causal convolution, dilated convolution, and residual connections. The temporal convolutional network uses causal convolution and dilated convolution to ensure the causality of the prediction, and enhances gradient flow through residual connections. The efficient channel attention mechanism extracts channel statistics through global average pooling, and then adaptively calculates the kernel size of the one-dimensional convolution based on the channel statistics. It uses one-dimensional convolution for cross-channel interaction, thereby adaptively calibrating the channel weights. The bidirectional long short-term memory network includes forward LSTM and backward LSTM. The bidirectional long short-term memory network captures global temporal dependencies through forward LSTM and backward LSTM.

[0009] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, the specific formula for the causal convolution is as follows: ; in, Indicates the index of causal convolution at time steps The extracted feature values, Indicates the input dataset. This indicates a causal convolution operation. For convolution kernel, For time step index, This represents the position index of the convolution kernel. Indicates the size of the convolution kernel; The formula for dilated convolution is: ; in, Indicates the index of the dilated convolution at time step. The extracted feature values, Indicates the expansion factor; The specific formula for the residual connection is: ; in, yes convolution, The residual function is defined by the following formula: ; in, This indicates a normalization operation. This represents the activation function.

[0010] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, the specific formula for global average pooling is as follows: ; in, For the first The time step The sample at the th The feature value of the i-th channel, the i-th For the first The sample at the th Global characteristics of each channel The total number of time steps for the input features; The expression for the one-dimensional convolution operation is: ; in, The weight parameters are those for a one-dimensional convolution. For the first The intermediate weight values ​​of each channel; The kernel size for one-dimensional convolution is calculated using the following formula: ; in, For the number of channels, and For hyperparameters; The specific formula for the adaptive calibration channel weights is as follows: ; ; in, For the first The time step The sample at the th The eigenvalues ​​after weighting each channel This represents the activation function. It is the first The final weight of each channel.

[0011] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, the bidirectional long short-term memory network is represented as follows: ; For each The forward LSTM representation at time step is: ; For each The inverse LSTM representation at time step 1 is: ; The output layer of the bidirectional long short-term memory network is: ; in, This indicates that bidirectional long short-term memory networks are in The merged hidden state vector at time step express The hidden state vector of the forward-biased LSTM at any given time. express The hidden state vector of the backward LSTM at each time step. This represents the activation function. This represents the weight matrix input to the hidden layer in the forward LSTM. express The input feature vector at time t, This represents the recurrent weight matrix of the hidden layer in a forward LSTM. Indicating a forward LSTM in The hidden state vector at time step 1. This represents the bias term of the forward LSTM hidden layer. This represents the weight matrix input to the hidden layer of the inverse LSTM. This represents the recurrent weight matrix of the hidden layer in the inverse LSTM. This represents the hidden state vector of the inverse LSTM at time t. This represents the bias term of the inverse LSTM hidden layer. express The output of the bidirectional LSTM at any given time. This represents the activation function of the output layer. This represents the weight matrix from the forward hidden state to the output layer. This represents the weight matrix from the hidden state to the output layer. This represents the bias term of the output layer.

[0012] In the aforementioned wind farm turbine speed anomaly early warning method based on fusion model, during fusion model training, the mean square error is used as the loss function, and the Adam optimizer is used to optimize the learning rate, training times, and time step hyperparameters until the fusion model converges and the optimal fusion model is obtained.

[0013] In the aforementioned wind farm turbine speed anomaly early warning method based on a fusion model, the specific formula for the loss function is as follows: ; in, For the sample size, For the first The true value of each sample For the fusion model to the first The predicted value for each sample.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The fusion model constructed in this invention, by integrating a temporal convolutional network, an efficient channel attention mechanism, and a bidirectional long short-term memory network architecture, can accurately predict the average generator speed of a wind turbine unit. By monitoring the prediction difference and the threshold of the predicted value, it can provide early warning of wind turbine overspeed, reducing the risk of overspeed accidents and maintenance costs of offshore generator units. Furthermore, the temporal convolutional network of this invention expands the receptive field while ensuring the causality of the prediction through causal convolution and dilated convolution, efficiently capturing long-term dependencies in the temporal sequence of speed. The efficient channel attention mechanism module dynamically calibrates channel weights by adaptively calculating the size of the one-dimensional convolutional kernel, highlighting key features and suppressing redundant information, without requiring dimensionality reduction and with low computational cost. The bidirectional long short-term memory network mines bidirectional temporal correlations through forward and backward networks, overcoming the limitations of single-direction modeling. This invention removes invalid features with zero or missing values ​​exceeding a threshold, deletes abnormal operation data of the unit, and combines the Pearson correlation coefficient method to screen core parameters that are highly correlated with the average generator speed. At the same time, it uses min-max normalization to eliminate dimensional differences, thereby filtering out low information density features and interference data. This significantly improves the signal-to-noise ratio of the input dataset, avoids model training bias caused by invalid data from the source, and provides a basic guarantee for subsequent prediction accuracy. Attached Figure Description

[0015] Figure 1 This is a flowchart of the wind turbine overspeed warning process of the present invention; Figure 2 This is a correlation analysis diagram of the present invention; Figure 3 This is a structural diagram of the fusion model of the present invention; Figure 4 This is a comparison chart of predicted and actual values ​​of overspeed monitoring data for Unit 8 (6.8MW) with a comprehensive dataset, as presented in this invention. Figure 5 This is a comparison chart of predicted and actual values ​​of overspeed monitoring for Unit 65# 8.3MW, based on a comprehensive dataset, according to the present invention. Figure 6 This is a radar chart showing the average generator speed of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0017] Example: A wind farm turbine speed anomaly early warning method based on a fusion model, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Collect multi-dimensional operation data of the SCADA system of a certain group of wind turbine units, and perform data preprocessing, normalization, correlation analysis and feature filtering on the operation data to obtain the input dataset; In this step, the multi-dimensional operational data of the SCADA system is collected by the SCADA system on the wind turbine. The input feature parameters in the operational data include average wind speed, average grid active power, average nacelle temperature, and average generator speed, which are the prediction target variables of the model. These are synchronously collected by the SCADA system as true values ​​for model training and error determination. The data preprocessing includes removing feature parameters whose proportion of zero and missing values ​​exceeds a threshold, and deleting data records of wind turbines in abnormal operating conditions. In this embodiment, the data preprocessing specifically involves: taking the 12# 8.3MW unit as the experimental object, binary features and categorical features, such as average unit yaw mode and average unit operating mode, should first be excluded. These data features have small variations and contain less information, so they need to be excluded first. At the same time, some feature parameters, although sensors are deployed, have not collected corresponding data and need to be excluded. Then, data with an average generator speed below 10 revolutions per minute are deleted. Data under abnormal operating conditions will significantly affect the training effect of the model and will also lead to abnormally large evaluation index values, failing to reflect the model training effect normally. At the same time, there is no risk of overspeed when the wind turbine is not running. The threshold for the proportion of zero and missing values ​​is set at 30%. That is, when the sum of the number of zero values ​​and the number of missing values ​​of a certain feature parameter accounts for more than 30% of the total data volume of that feature, it is determined to be an invalid feature and removed.

[0018] The normalization process applies a minimum-maximum normalization method to the remaining running data, transforming the feature dataset of the running data through a linear transformation. Any original feature parameters in The mapping is over the interval [0,1], and the specific formula is as follows: ; in, This represents the feature parameters after normalization. Representing the feature dataset The minimum value of all characteristic parameters in. Representing the feature dataset The maximum value of all feature parameters in the data; like Figure 2 As shown, the correlation analysis uses the Pearson correlation coefficient method to calculate the characteristic parameters. By analyzing the correlation coefficients between the parameters, we can select the feature parameters that are highly correlated with the average generator speed, and obtain the input dataset. .

[0019] This invention removes invalid features with zero or missing values ​​exceeding a threshold, deletes abnormal operation data of the unit, and combines the Pearson correlation coefficient method to screen core parameters that are highly correlated with the average generator speed. At the same time, it uses min-max normalization to eliminate dimensional differences, thereby filtering out low information density features and interference data. This significantly improves the signal-to-noise ratio of the input dataset, avoids model training bias caused by invalid data from the source, and provides a basic guarantee for subsequent prediction accuracy.

[0020] Step 2: Construct a fusion prediction model composed of a temporal convolutional network, an efficient channel attention mechanism, and a bidirectional long short-term memory network. Input the input dataset into the fusion model for training. Optimize the hyperparameters to achieve the best prediction effect and obtain the optimal fusion model. In this step, an input time series sequence is given. And predict using sequence modeling networks The corresponding output time sequence at time 1 The specific formula is as follows: ; in, This represents a sequence modeling network.

[0021] The goal of learning in the sequence modeling setting is to find a sequence modeling network. When the input sequence and the true output sequence follow a specific distribution, it is possible to minimize a certain expected loss between the actual output and the prediction. The sequence and output are extracted according to a certain distribution.

[0022] like Figure 3 The diagram shows the structure of the entire fusion model. TCN employs multi-layer dilated convolutions to expand the receptive field while maintaining causality, enabling efficient modeling of long-term dependencies in time series. BiLSTM enhances global feature representation by mining forward and backward correlation features of the sequences. The ECA module enables dynamic weight allocation along the channel dimension, highlighting key features and suppressing redundant information. The fusion model is represented as follows: ; ; ; ; in, Indicates the input dataset. Represents a temporal convolutional network. Temporal features extracted by a temporal convolutional network. This indicates an efficient channel attention mechanism. Features weighted for efficient channel attention mechanisms It is a bidirectional long short-term memory network. The hidden state features output by BiLSTM. It is a fully connected layer. These are the predicted values ​​from the fusion model.

[0023] Temporal Convolutional Networks (TCNs) include causal convolution, dilated convolution, and residual connections. TCNs use causal and dilated convolutions to ensure the causality of predictions and residual connections to enhance gradient flow. TCNs are based on two principles: the network must produce an output of the same length as the input, and there must be no leakage of information from the future to the past. To achieve the former, TCNs employ a one-dimensional fully convolutional network architecture, using zero-padding to maintain the same length for each hidden layer as the input layer; to achieve the latter, causal convolution is used, i.e., in time... The output is only related to the previous layer. Convolution is performed on elements at time points earlier than the given time, therefore TCN can be defined as a combination of 1DFCN and causal convolution. Since causal convolution can only recall historical data linearly related to network depth, dilated convolution is used to achieve an exponentially large receptive field. For inputs of one-dimensional sequences... and a filter The specific formula for the causal convolution is as follows: ; in, Indicates the index of causal convolution at time steps The extracted feature values, Indicates the input dataset. This indicates a causal convolution operation. For convolution kernel, For time step index, This represents the position index of the convolution kernel. Indicates the size of the convolution kernel; The formula for dilated convolution is: ; in, Indicates the index of the dilated convolution at time step. The extracted feature values, Indicates the expansion factor; The specific formula for the residual connection is: ; in, It's convolution. The residual function is defined by the following formula: ; in, This indicates a normalization operation. This represents the activation function.

[0024] In this embodiment, the invention uses a larger inflation factor so that the output of the top layer can represent a wider range of inputs, thereby effectively expanding the receptive field of the convolutional network. The residual block of the temporal convolutional network contains a branch that leads to a series of transformations, effectively allowing the layer to learn modifications to the identity mapping rather than the entire transformation.

[0025] In this step, the fusion model extracts channel statistics through global average pooling, and then adaptively calculates the kernel size of the one-dimensional convolution based on the channel statistics. One-dimensional convolution is used for cross-channel interaction, thereby adaptively calibrating the channel weights. This avoids the dimensionality reduction operation in traditional attention mechanisms, achieving adaptive calibration of channel features with extremely low computational cost while maintaining the direct correspondence between channels and weights. Specifically, the ECA module first adaptively calculates the kernel size of the one-dimensional convolution based on the number of channels. The calculation formula is as follows: ; in, For the number of channels, and As a hyperparameter, obtain the kernel size. The ECA module then applies one-dimensional convolutions to the input features to learn the importance of each channel relative to the other channels.

[0026] The specific formula for global average pooling is: ; in, For the first The time step The sample at the th The feature value of the i-th channel, the i-th For the first The sample at the th Global characteristics of each channel This represents the total number of time steps for the input features.

[0027] The expression for the one-dimensional convolution operation in the efficient channel attention mechanism is: ; in, The weight parameters are those for a one-dimensional convolution. For the first The intermediate weight values ​​of each channel; The specific formula for the adaptive calibration channel weights is as follows: ; ; in, For the first The time step The sample at the th The eigenvalues ​​after weighting each channel This represents the activation function. It is the first The final weight of each channel.

[0028] In temporal convolutional networks, the Bidirectional Long Short-Term Memory (BiLSTM) network includes a forward LSTM and a backward LSTM. The BiLSTM network captures global temporal dependencies through these forward and backward LSTMs. The BiLSTM network is represented as follows: ; For each The forward LSTM representation at time step is: ; For each The inverse LSTM representation at time step 1 is: ; The output layer of the bidirectional long short-term memory network is: ; in, This indicates that bidirectional long short-term memory networks are in The merged hidden state vector at time step express The hidden state vector of the forward-biased LSTM at any given time. express The hidden state vector of the backward LSTM at each time step. This represents the activation function. This represents the weight matrix input to the hidden layer in the forward LSTM. express The input feature vector at time t, This represents the recurrent weight matrix of the hidden layer in a forward LSTM. Indicating a forward LSTM in The hidden state vector at time step 1. This represents the bias term of the forward LSTM hidden layer. This represents the weight matrix input to the hidden layer of the inverse LSTM. This represents the recurrent weight matrix of the hidden layer in the inverse LSTM. Indicates that the inverse LSTM is in The hidden state vector at time step 1. This represents the bias term of the inverse LSTM hidden layer. express The output of the bidirectional LSTM at any given time. This represents the activation function of the output layer. This represents the weight matrix from the forward hidden state to the output layer. This represents the weight matrix from the hidden state to the output layer. This represents the bias term of the output layer.

[0029] BiLSTM uses a bidirectional network to recursively feed back the hidden states of the past and future, thereby further utilizing the intrinsic relationship between the current payload data and the payload data of the past and future time points, and further improving the prediction accuracy and feature data utilization of the fusion model.

[0030] Step 3: The average generator speed of the target wind turbine is predicted using the optimal fusion model to obtain the speed prediction result. The deviation between the speed prediction result and the actual speed result is calculated. The deviation is judged based on a preset threshold. When the abnormal signal accumulates to exceed the set condition, an early warning is triggered to complete the abnormal speed detection of the wind turbine.

[0031] During the training of the fusion model, the TCN-ECA-BiLSTM is optimized with respect to three hyperparameters: learning rate, number of training epochs, and timestep. The mean squared error is used as the loss function during training, and the specific formula for the loss function is as follows: ; in, For the sample size, For the first The true value of each sample For the fusion model to the first The predicted value for each sample.

[0032] During the training process of the fusion model, TCN performs long-term feature extraction, ECA adaptively weights key channels, and BiLSTM fuses the time-series information before and after the training to obtain the fusion model with the best prediction performance, which is then used to predict the average generator speed. The model's prediction performance is then evaluated using the root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

[0033] In this embodiment, the preset threshold is determined based on historical anomaly data statistics. The setting condition is 5 consecutive anomaly signals within 3 time steps or 5 minutes (anomaly signal refers to a signal where the error between the predicted value and the true value exceeds the preset threshold). The hyperparameter optimization range is: learning rate 1e-5~1e-3, training times 50-200 rounds, and time steps 10-30.

[0034] In the dataset used in this study, most wind turbines only provide limited characteristic parameters such as average wind direction, average grid active power, average atmospheric humidity, and average nacelle temperature. Furthermore, the data collection period for wind turbine operation is as long as one year, and the data has a strong correlation with time. Therefore, months were additionally introduced as a time dimension parameter into the fusion model training. To avoid overfitting, the dataset was divided into a training set (first 8 months), a validation set (9th-10th months), and a test set (last 2 months) according to time series, without random shuffling (to preserve time series correlation). It should be noted that most of the above characteristic parameters have low correlation with average generator speed, which poses a challenge to accurately estimating the current real-time generator speed. The prediction accuracy of average generator speed for different units is shown in Table 1.

[0035] Table 1

[0036] As can be seen from the quantitative results in Table 1, the fusion model exhibits excellent performance across different turbine models and datasets: For both the 6.8MW and 8.3MW turbine models, the R² (decision coefficient) values ​​of the prediction results are close to 1, indicating a higher model fit. Furthermore, the RMSE (root mean square error) and MAE (mean absolute error) are close to 0, indicating a smaller deviation between the predicted and actual values ​​and higher model accuracy. This fully validates the model's stability and prediction accuracy. Based on this, a dedicated overspeed warning model was further constructed for the complete datasets of the 8# 6.8MW and 65# 8.3MW wind turbine units. The warning effect is as follows: Figure 4 , Figure 5 As shown.

[0037] To avoid overfitting due to excessive reliance on temporal sequence patterns during model training, the dataset was randomly shuffled before training. Figure 4 , Figure 5 The visualization results show that the predicted curve of the average generator speed coincides extremely well with the actual observed curve—this phenomenon fully demonstrates that the feature parameters selected through the previous correlation analysis play a crucial supporting role in the model's prediction accuracy. Further verification of the predictive performance of the overflight monitoring model was conducted using quantitative evaluation indicators: the R² of the 6.8MW unit #8 reached 0.9940, RMSE was 0.0210, and MAE was 0.0122; the R² of the 8.3MW unit #65 was even higher at 0.9984, with an RMSE as low as 0.0126 and MAE as low as 0.0094. Both units' indicators demonstrate the model's high predictive reliability.

[0038] To fully verify the superiority of the fusion model, its performance on the prediction tasks of two core features—"average generator speed" and "average wind speed"—was compared with four commonly used classic models in the field: Bidirectional Long Short-Term Memory (BiLSTM), TCN-ECA, Temporal Convolutional Network (TCN), and Convolutional Neural Network-BiLSTM-ATT with an attention mechanism. The comparison results show that the fusion model significantly outperforms all the comparison models in the prediction of the above two core features. To present the performance differences of each model more intuitively and comprehensively, Figure 6 The radar charts visualize the three core evaluation metrics (R², RMSE, and MAE) for the "average generator speed" prediction task, clearly reflecting the advantages and disadvantages of different models in each metric dimension, and further highlighting the comprehensive performance advantages of the fusion model.

[0039] In summary, the fusion model constructed in this invention, by integrating a temporal convolutional network, an efficient channel attention mechanism, and a bidirectional long short-term memory network architecture, can accurately predict the average generator speed of a wind turbine unit. By monitoring the prediction difference and the threshold of the prediction value, it can complete the wind turbine unit overspeed warning, thereby reducing the potential risks and maintenance costs of offshore wind turbine unit overspeed monitoring.

Claims

1. A wind farm turbine speed anomaly early warning method based on a fusion model, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional operation data of the SCADA system of a certain group of wind turbine units, and perform data preprocessing, normalization, correlation analysis and feature filtering on the operation data to obtain the input dataset; Step 2: Construct a fusion prediction model composed of a temporal convolutional network, an efficient channel attention mechanism, and a bidirectional long short-term memory network. Input the input dataset into the fusion model for training. Optimize the hyperparameters to achieve the best prediction effect and obtain the optimal fusion model. Step 3: The average generator speed of the target wind turbine is predicted using the optimal fusion model to obtain the speed prediction result. The deviation between the speed prediction result and the actual speed result is calculated. The deviation is judged based on a preset threshold. When the abnormal signal accumulates to exceed the set condition, an early warning is triggered to complete the wind turbine speed abnormality early warning.

2. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: In step one, the characteristic parameters in the multi-dimensional operating data of the SCADA system include average generator speed, average wind speed, average grid active power, and average nacelle temperature.

3. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: The data preprocessing includes removing feature parameters in the operating data whose proportion of zero and missing values ​​exceeds a threshold, and deleting data records in the operating data where the wind turbine is in an abnormal operating state. The normalization process applies a minimum-maximum normalization method to the remaining running data, transforming the feature dataset of the running data through a linear transformation. Any original feature parameters in The mapping is over the interval [0,1], and the specific formula is as follows: ； in, This represents the feature parameters after normalization. Representing the feature dataset The minimum value of all characteristic parameters in. Representing the feature dataset The maximum value of all feature parameters in the data; The correlation analysis uses the Pearson correlation coefficient method to calculate the characteristic parameters. By analyzing the correlation coefficients between the parameters, we can select the feature parameters that are highly correlated with the average generator speed, and obtain the input dataset. .

4. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: In step two, the fusion model is represented as follows: ； ； ； ； in, Indicates the input dataset. Represents a temporal convolutional network. Temporal features extracted by a temporal convolutional network. This indicates an efficient channel attention mechanism. Features weighted for efficient channel attention mechanisms It is a bidirectional long short-term memory network. The hidden state features output by BiLSTM. It is a fully connected layer. These are the predicted values ​​from the fusion model.

5. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: The temporal convolutional network includes causal convolution, dilated convolution, and residual connections. Causal and dilated convolutions ensure the causality of predictions, while residual connections enhance gradient flow. The efficient channel attention mechanism extracts channel statistics through global average pooling, then adaptively calculates the kernel size of the one-dimensional convolution based on these statistics, using one-dimensional convolution for cross-channel interaction to adaptively calibrate channel weights. The bidirectional long short-term memory network includes forward LSTM and backward LSTM, capturing global temporal dependencies.

6. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 5, characterized in that: The specific formula for the causal convolution is as follows: ； in, Indicates the index of causal convolution at time steps The extracted feature values, Indicates the input dataset. This indicates a causal convolution operation. For convolution kernel, For time step index, This represents the position index of the convolution kernel. Indicates the size of the convolution kernel; The formula for dilated convolution is: ； in, Indicates the index of the dilated convolution at time step. The extracted feature values, Indicates the expansion factor; The specific formula for the residual connection is: ； in, yes convolution, The residual function is defined by the following formula: ； in, This indicates a normalization operation. This represents the activation function.

7. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 5, characterized in that: The specific formula for global average pooling is as follows: ； in, For the first The time step The sample at the th The feature value of the i-th channel, the i-th For the first The sample at the th Global characteristics of each channel The total number of time steps for the input features; The expression for the one-dimensional convolution operation is: ； in, The weight parameters are those for a one-dimensional convolution. For the first The intermediate weight values ​​of each channel; The kernel size for one-dimensional convolution is calculated using the following formula: ； in, For the number of channels, and For hyperparameters; The specific formula for the adaptive calibration channel weights is as follows: ； ； in, For the first The time step The sample at the th The eigenvalues ​​after weighting each channel This represents the activation function. It is the first The final weight of each channel.

8. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 5, characterized in that: The bidirectional long short-term memory network is represented as follows: ； For each The forward LSTM representation at time step is: ； For each The inverse LSTM representation at time step 1 is: ； The output layer of the bidirectional long short-term memory network is: ； in, This indicates that bidirectional long short-term memory networks are in The merged hidden state vector at time step express The hidden state vector of the forward-biased LSTM at any given time. express The hidden state vector of the backward LSTM at each time step. This represents the activation function. This represents the weight matrix input to the hidden layer in the forward LSTM. express The input feature vector at time t, This represents the recurrent weight matrix of the hidden layer in a forward LSTM. Indicating a forward LSTM in The hidden state vector at time step 1. This represents the bias term of the forward LSTM hidden layer. This represents the weight matrix input to the hidden layer of the inverse LSTM. This represents the recurrent weight matrix of the hidden layer in the inverse LSTM. Indicates that the inverse LSTM is in The hidden state vector at time step 1. This represents the bias term of the inverse LSTM hidden layer. express The output of the bidirectional LSTM at any given time. This represents the activation function of the output layer. This represents the weight matrix from the forward hidden state to the output layer. This represents the weight matrix from the hidden state to the output layer. This represents the bias term of the output layer.

9. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: During the training of the fusion model, the mean squared error is used as the loss function, and the Adam optimizer is used to optimize the learning rate, number of training iterations, and time step hyperparameters until the fusion model converges, thus obtaining the optimal fusion model.

10. The wind farm turbine speed anomaly early warning method based on a fusion model according to claim 1, characterized in that: The specific formula for the loss function is as follows: ； in, For the sample size, For the first The true value of each sample For the fusion model to the first The predicted value for each sample.

Images