Wind turbine generator yaw system fault early warning method based on prediction deviation

Abstract

The invention belongs to the field of wind turbine generator state monitoring and fault early warning, and relates to a wind turbine generator yaw system fault early warning method based on prediction deviation, and the method comprises the steps: collecting historical operation data of a fan SCADA system, and carrying out the preprocessing of the historical operation data; feature parameters with high correlation are extracted as input quantities of a prediction model, and a grey wolf algorithm is improved to optimize a long-short-term memory neural network model to predict a yaw angle; calculating a yaw angle prediction deviation value by combining a sliding window and a probability density function, and determining an early warning threshold value; and according to the real-time yaw angle recorded by the SCADA and the corresponding prediction data, whether the prediction deviation exceeds a threshold value is calculated so as to judge whether fault early warning is carried out or not. According to the method, the advantages of the improved grey wolf algorithm and the long-short-term memory network in the optimization capability and the time sequence data processing capability are combined, so that the prediction accuracy is improved, and the fault early warning accuracy and reliability are further improved.

Description

Technical Field

[0001] The invention belongs to the field of wind turbine status monitoring and fault early warning, and relates to a wind turbine yaw system fault early warning method based on prediction deviation. Background Art

[0002] With the rapid development of wind power, the number of wind turbines installed is increasing, which makes the maintenance pressure of wind turbines increasingly greater. Real-time monitoring of wind turbine operation status and timely and accurate warning of faults are important tasks to reduce downtime, reduce operation and maintenance costs and ensure the safety of the unit. Among the main components of wind turbines, the failure frequency of the yaw system and the downtime caused by the failure account for too high a proportion, which has become an important consideration for the safe and reliable operation of wind turbines. Therefore, establishing an effective fault warning method can help the unit improve safety, stability and economy.

[0003] Modern wind farms are all equipped with SCADA systems. For wind turbine fault warning, the cost of using SCADA data is relatively low, and the use of SCADA data for fault analysis has attracted more and more attention from the academic community. However, wind turbines are generally distributed more dispersedly and have a harsh operating environment. They are affected by random factors such as external wind speed changes and internal mechanical instability, resulting in a wide distribution of SCADA data sampling values, inconsistent sampling frequencies, invalidity, and missing, and interference data such as null values ​​and abnormal values ​​that cannot reflect the true state of the wind turbine. These distorted data will seriously affect the accuracy of data analysis. It is difficult to directly find physical information that reflects the specific state of the yaw system with a large number of characteristic parameters of the SCADA system, and it is impossible to conduct qualitative analysis. In order to improve the efficiency and accuracy of the model operation, the best performance model should be obtained with fewer input parameters. The correlation with the target parameters should be increased as much as possible, and the weight of the relevant parameters should be increased.

[0004] Since deep learning algorithms are suitable for processing large-scale, high-dimensional data, have strong pattern recognition capabilities, and are highly adaptable and scalable, they are widely used in wind turbine fault warning. LSTM has good processing capabilities for long time series, and LSTM units can store useful data for a long time. Wind turbine SCADA data is diverse and the data volume is huge, and good results can be achieved by using LSTM neural network processing. Traditional LSTM has the disadvantages of high computational complexity and possible convergence to local values. The improved gray wolf algorithm has outstanding optimization capabilities and has the advantages of global convergence and fast convergence speed. Therefore, the improved gray wolf algorithm is applied to the parameters of the improved LSTM. Summary of the invention

[0005] In view of this, the present invention provides a wind turbine fault early warning method and device based on abnormal deviation detection.

[0006] The present invention discloses a wind turbine yaw system fault early warning method based on prediction deviation, which comprises:

[0007] Step 1: Preprocessing of wind turbine SCADA data and selection of characteristic parameters; preprocessing includes the processing of null values ​​and outliers, and characteristic parameter selection uses Pearson linear correlation to sort the weights of SCADA parameters and select SCADA parameters that can better reflect the operating status characteristics of the yaw system.

[0008] Step 2: Use the characteristic parameter variables screened in step 1 as input and the yaw angle as output to establish a multi-input single-output regression prediction model.

[0009] Step 3: Input the test set into the trained IGWO-LSTM regression prediction model, calculate the predicted deviation between the predicted value and the true value of the output yaw angle, and then fit the probability density function of the predicted deviation. Set a confidence interval to calibrate the normal range of the predicted deviation, and define the confidence interval as the upper and lower boundaries. The values ​​of the upper and lower boundaries are the yaw system fault warning thresholds.

[0010] Step 4: Obtain online real-time monitoring multivariate time data from the wind turbine, input it into the trained IGWO-LSTM regression prediction model after preprocessing and feature selection, calculate the prediction deviation between the predicted value and the true value in combination with the time sliding window, and compare it with the pre-obtained threshold; when the deviation value is greater than the threshold, a fault warning is issued for the wind turbine component.

[0011] A further improvement of the technical solution of the present invention is that step 1 includes the following steps:

[0012] The min-max standardization method is adopted to normalize the SCADA data and convert or unify the data format into a form suitable for modeling.

[0013] The outlier detection algorithm is used to detect outliers in the original SCADA data and remove outliers that are significantly different from other data points.

[0014] Screen the variable parameters related to the operation status of the wind turbine yaw system, including the angle of the twist and hold, the temperature outside the cabin, the temperature inside the cabin, the active power of the outer side, the active power of the generator, the wind direction, the wind speed, the yaw motor speed, the generator speed, the yaw angle, etc. Use the Pearson correlation coefficient method to perform correlation analysis on these parameters and select the SCADA parameters with larger weight values.

[0015] A further improvement of the technical solution of the present invention is that in step 2, the following steps are included: establishing an LSTM time series prediction model based on an improved gray wolf optimization algorithm, using the improved gray wolf algorithm to optimize the LSTM neural network hyperparameters, reducing the generalization error through hyperparameter tuning, and ensuring that the model can converge to the global optimum.

[0016] After multiple debugging, the specific structure of the network model is set to: one layer of sequence input, two LSTM layers and one fully connected layer. Setting two Dropout layers with a dropout rate of 21% can alleviate the overfitting phenomenon, and the last layer is the regression output.

[0017] The first 80% of the data is used as the training set, the last 10% as the test set, and the middle 10% as the validation set.

[0018] The further improvement of the technical solution of the present invention is that: in step 3, the following steps are included: in order to reduce the probability of false alarm of the early warning, based on prior knowledge, the confidence interval is set to 99%, then in the probability density curve, the value corresponding to 0.5% is the lower boundary, and the value corresponding to 99.5% is the upper boundary. First, the mean and standard deviation of all deviation group data in each sliding window are calculated. Then the probability density function of the predicted deviation data is calculated, and the probability density curve of the predicted deviation is obtained to determine the upper and lower boundary values, thereby determining the threshold of the early warning model.

[0019] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0020] 1. In order to improve the efficiency and accuracy of the model operation, the model with the best performance should be obtained with fewer input parameters. The correlation with the target parameters should be increased as much as possible, and the weight of the relevant parameters should be increased. The Pearson linear correlation coefficient is the most commonly used linear correlation coefficient, which can intuitively judge the importance of the parameters. This advantage can better select the characteristic parameters of the same model under actual operation conditions, and better apply to the units of the same wind farm.

[0021] 2. Apply the improved Gray Wolf algorithm to the parameters of the improved LSTM, and apply the LSTM time series prediction model based on the improved Gray Wolf optimization algorithm to the prediction field of time series. This model not only absorbs the advantages of the long short-term memory network that is good at processing time series data, but also uses the improved Gray Wolf algorithm to determine the network parameters to ensure that the model can converge to the global optimum. The generalization error is reduced by hyperparameter tuning, thereby improving the accuracy of yaw angle prediction and the reliability of fault warning.

[0022] 3. Combined with the sliding window to process the prediction deviation, the fault warning of the unit can be realized, which can effectively reduce the number of false alarms. Description of the drawings:

[0023] Figure 1 It is a fault warning flow chart of the present invention.

[0024] Figure 2 is the probability density plot of the forecast deviation.

[0025] Figure 3 This is the fault warning experiment result diagram. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below in conjunction with embodiments:

[0027] Figure 1 The present invention is a flow chart of a fault warning method for a yaw system of a wind turbine generator set based on prediction deviation. The method includes SCADA historical data preprocessing, feature parameter selection, establishing an IGWO-LSTM prediction model, substituting the parameter training model, calculating the warning threshold, bringing in SCADA real-time data to calculate the prediction deviation between the predicted value and the true value, and comparing with the threshold to determine whether a fault warning is issued.

[0028] The min-max standardization method was used to normalize the SCADA data, and the data format was converted or unified into a form suitable for modeling. The outlier factor detection algorithm was used to detect outliers in the original SCADA data, and outliers that were significantly different from other data points were removed. The Pearson correlation coefficient method was used for feature selection, and the top 6 SCADA parameters with the highest weight values ​​were selected, namely, the angle of the twist and hold, the temperature outside the cabin, the temperature inside the cabin, the active power of the outside, the active power of the generator, and the wind direction.

[0029] An LSTM time series prediction model based on the improved gray wolf optimization algorithm is established. The improved gray wolf algorithm is used to optimize the LSTM neural network hyperparameters. The generalization error is reduced by hyperparameter tuning to ensure that the model can converge to the global optimum. After multiple debugging, the specific structure of the network model is set to: one layer of sequence input, 2 LSTM layers and 1 fully connected layer. Setting 2 Dropout layers with a dropout rate of 21% can alleviate the overfitting phenomenon, and the last layer is the regression output. The first 80% of the data is used as the training set, the last 10% as the test set, and the middle 10% as the validation set. The test set is input into the trained IGWO-LSTM regression prediction model to calculate the predicted value of the output yaw angle and the predicted deviation of the true value.

[0030] Figure 2The probability density curve of the forecast deviation calculated using historical data is shown. To reduce the probability of false alarms, the confidence interval β is set to 99% based on prior knowledge. Then, in the probability density curve, the value corresponding to 0.5% is the lower boundary, and the value corresponding to 99.5% is the upper boundary. First, the mean of all data in each deviation group is calculated. μ and standard deviation σ .

[0031]

[0032] Then calculate the probability density function of the prediction deviation data f ( x ), the formula is as follows:

[0033]

[0034] The upper boundary value was calculated to be 3.72 and the lower boundary value was -3.25, thereby determining the threshold of the early warning model.

[0035] In order to verify the accuracy of the IGWO-LSTM prediction model for fault warning, the SCADA historical fault data of the same type of unit in the above wind farm was used for experiments. The unit had a "crack in the yaw gear" fault alarm signal at 15:45 on April 20, 2017. 1000 operating data within a period of time before the fault occurred were selected for analysis, and the size of the sliding window was set to 20.

[0036] Fault warning results such as Figure 3 As shown in the figure, the predicted deviation of the yaw angle exceeded the warning threshold at the 483rd point, and the warning was issued 527 points in advance before the yaw gear failure occurred, which discovered the potential fault early and avoided the abnormal shutdown caused by the yaw gear failure. According to the calculation, the corresponding time was 6:58 on April 20, 2017, which was 8 hours and 47 minutes earlier than the actual fault time.

Claims

1. A wind turbine yaw system fault warning method based on prediction deviation, characterized in that: The steps include: Step 1: Preprocessing of wind turbine SCADA data and selection of characteristic parameters; preprocessing includes the processing of null values ​​and outliers, and characteristic parameter selection uses Pearson linear correlation to sort the weights of SCADA parameters and select SCADA parameters that can better reflect the operating status characteristics of the yaw system. Step 2: Use the characteristic parameter variables screened in step 1 as input and the yaw angle as output to establish a multi-input single-output regression prediction model. Step 3: Input the test set into the trained IGWO-LSTM regression prediction model, calculate the predicted deviation between the predicted value and the true value of the output yaw angle, and then fit the probability density function of the predicted deviation. Set a confidence interval to calibrate the normal range of the predicted deviation, and define the confidence interval as the upper and lower boundaries. The values ​​of the upper and lower boundaries are the yaw system fault warning thresholds. Step 4: Obtain online real-time monitoring multivariate time data from the wind turbine, input it into the trained IGWO-LSTM regression prediction model after preprocessing and feature selection, calculate the prediction deviation between the predicted value and the true value in combination with the time sliding window, and compare it with the pre-obtained threshold; when the deviation value is greater than the threshold, a fault warning is issued for the wind turbine component.

2. According to claim 1, a wind turbine yaw system fault warning method based on prediction deviation is characterized in that Step 1 includes the following steps: The min-max standardization method is adopted to normalize the SCADA data and convert or unify the data format into a form suitable for modeling. The outlier detection algorithm is used to detect outliers in the original SCADA data and remove outliers that are significantly different from other data points. Screen the variable parameters related to the operation status of the wind turbine yaw system, including the angle of the twist and hold, the temperature outside the cabin, the temperature inside the cabin, the active power of the outer side, the active power of the generator, the wind direction, the wind speed, the yaw motor speed, the generator speed, the yaw angle, etc. Use the Pearson correlation coefficient method to perform correlation analysis on these parameters and select the SCADA parameters with larger weight values.

3. A wind turbine yaw system fault warning method based on prediction deviation according to claim 2, characterized in that Step 2 includes the following steps: An LSTM time series prediction model based on the improved gray wolf optimization algorithm is established. The improved gray wolf algorithm is used to optimize the LSTM neural network hyperparameters. The generalization error is reduced by hyperparameter tuning to ensure that the model can converge to the global optimum. After multiple debugging, the specific structure of the network model is set to: one layer of sequence input, two LSTM layers and one fully connected layer. Setting two Dropout layers with a dropout rate of 21% can alleviate the overfitting phenomenon, and the last layer is the regression output. The first 80% of the data is used as the training set, the last 10% as the test set, and the middle 10% as the validation set.

4. A wind turbine yaw system fault warning method based on prediction deviation according to claim 3, characterized in that Step 3 includes the following steps: In order to reduce the probability of false alarms, based on prior knowledge, the confidence interval is set to 99%. Then, in the probability density curve, the value corresponding to 0.5% is the lower boundary, and the value corresponding to 99.5% is the upper boundary. First, the mean and standard deviation of all deviation group data in each sliding window are calculated. Then the probability density function of the predicted deviation data is calculated, and the probability density curve of the predicted deviation is obtained to determine the upper and lower boundary values, thereby determining the threshold of the warning model.

Images