Wind turbine generator group galloping early warning method in same offshore wind field
By building the XGBoost model and using the GAIN network for data filling, combined with transfer learning technology, the speed warning of offshore wind turbines is achieved, and the problems of low monitoring accuracy and high cost in the existing technology are solved, and the accuracy and efficiency of early warning are improved.
Patent Information
- Application Number
- CN202411795437.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-03-14
AI Technical Summary
In the prior art, the accuracy of the speed warning monitoring of offshore wind turbines is not high and the cost is too high, making it difficult to effectively prevent and respond to speed accidents.
The XGBoost model is used to predict the operating data of offshore wind turbines, fill in the missing data through the GAIN network, and transfer learning is used to transfer the trained model to other wind turbines in the same wind farm to achieve a speed warning for the entire wind farm.
It improves the accuracy of speed warning and monitoring, reduces the hidden dangers and early warning costs of offshore wind turbines, and enhances the safety and operational efficiency of wind turbines.
Smart Images
Figure CN119622213A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fault diagnosis of offshore wind turbine generator system, and specifically relates to a method for early warning of wind turbine group runaway in the same offshore wind farm. Background Art
[0002] In recent years, wind energy has been increasingly valued due to its advantages of being clean, pollution-free, renewable and inexhaustible, and is becoming the main part of renewable energy. As a major energy consumer, China is optimizing its energy structure and upgrading its energy industry in accordance with the "dual carbon" policy. Various measures are being actively taken to promote the development of the wind power industry. Among them, the development of onshore wind power is relatively early and has been relatively saturated. China has a long coastline and is located in the Asian monsoon region, with abundant offshore wind energy resources. At the same time, compared with onshore wind power, offshore wind power does not need to occupy a large area of land, and the noise far away from the coast has less impact on residents. The sea breeze is stronger and more stable than the land breeze. At the same time, offshore wind power can directly provide electricity to coastal cities to ensure the safety of energy supply. In addition, offshore wind power can also optimize my country's energy structure and promote the upgrading of the energy industry. Therefore, the development of offshore wind power is a new development direction for renewable energy.
[0003] Since offshore wind turbines are often exposed to harsh environments, the maintenance cost of offshore wind turbines is much higher than that of onshore wind turbines. Wind turbines are more susceptible to corrosion and wear, and have higher hidden dangers of failure. At the same time, most offshore wind turbines are located in remote areas. Some offshore wind farms are even located tens of kilometers away from the coast, which is not conducive to transportation and maintenance testing. Once a runaway accident occurs, it is extremely difficult to repair and protect, and the maintenance cost is high. However, a runaway accident is a serious safety accident for wind turbines. A runaway accident is usually caused by aging and wear of various components of the wind turbine, excessive wind speed, etc., which causes the braking system or pitch system of the wind turbine to fail. The pitch system cannot control the blades to reduce the speed in time, causing the impeller speed to exceed the rated speed, making the unit out of control. Moreover, once a runaway accident occurs, it is difficult to rescue in time, which will destroy various components of the wind turbine. In severe cases, it may even cause fires, unit tower collapse and other serious consequences, which are easy to cause casualties and property losses. In the prior art, the runaway warning monitoring for offshore wind turbines has the problems of low accuracy and high monitoring cost. Summary of the invention
[0004] The purpose of the present invention is to provide a method for early warning of runaway wind turbines in a wind farm at sea. The present invention can improve the accuracy of early warning monitoring of runaway wind turbines and reduce the hidden dangers of monitoring runaway wind turbines at sea and the early warning cost.
[0005] The technical solution of the present invention is a method for early warning of wind turbine group runaway in the same offshore wind farm, comprising the following steps:
[0006] Step 1: Obtain the operating data of a group of offshore wind turbines, perform data preprocessing, correlation analysis and feature screening on the operating data, and fill the data through the GAIN network to obtain a data set;
[0007] Step 2: Build an XGBoost model and input the data set into the XGBoost model for training until the model prediction effect reaches the best and the optimal XGBoost model is obtained;
[0008] Step 3: using the optimal XGBoost model, predict the average generator speed of the offshore wind turbines, and monitor the difference between the predicted value and the true value to complete the runaway warning of the offshore wind turbines in the group;
[0009] Step 4: Use the GAIN network to fill in the data of other offshore wind turbines with missing data, and then use transfer learning to migrate the trained optimal XGBoost model to other offshore wind turbines in the same wind farm, so as to achieve runaway warning for all offshore wind turbines in the entire wind farm.
[0010] In the aforementioned method for early warning of runaway wind turbines in a wind farm at sea, in step 1, the operating data of the offshore wind turbines are collected by a SCADA system on the offshore wind turbines.
[0011] In the aforementioned method for early warning of wind turbine group runaway in the same offshore wind farm, the data preprocessing includes eliminating the maximum value features and category features in the offshore wind turbine operation data, removing feature parameters with a large proportion of zero values and missing values, and deleting data under abnormal working conditions of the wind turbine.
[0012] In the aforementioned method for early warning of group runaway of wind turbines in the same offshore wind farm, the correlation analysis specifically uses the Pearson correlation coefficient method to calculate the correlation coefficient between each pair of feature parameters for preliminary screening; the feature screening is to screen the remaining feature parameters according to the feature importance index.
[0013] In the aforementioned method for early warning of wind turbine group runaway in the same offshore wind farm, in step 2, the process of constructing the XGBoost model is:
[0014] Given a model with n samples and m features:
[0015] Q=(x i ,y i )(|Q|=n,x i ∈R m ,y i ∈R);
[0016] The predicted value function in the prediction process is as follows:
[0017]
[0018] Where i represents the i-th sample in the training set; is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree;
[0019] The objective function is defined as follows:
[0020]
[0021] Where Obj is the objective function; i is the i-th sample in the training set; y i is the true value of the i-th sample; is the i-th sample x i The predicted value of is the loss function; Ω(f k ) is the regularization term;
[0022] The loss function is defined as follows:
[0023]
[0024] The regularization term is defined as follows:
[0025]
[0026] Where T is the number of leaf nodes in the decision tree; ω represents the score on each leaf node; γ and λ represent regularization parameters;
[0027] The XGBoos model function obtained by the t-th training is defined as follows:
[0028]
[0029] Among them, Obj (t) is the objective function obtained from the tth learning; Ω(f t ) is the regularization term of the XGBoos model function during the t-th learning; y i is the true value of the i-th sample; For the tth learning, the i-th sample x i The predicted value of; C is a constant;
[0030] The gain function formula is defined as follows:
[0031] Gain=Obj (L+R) -(Obj L +Obj L );
[0032] Where, Gain is the gain function; Obj L The objective function for predicting the left leaf node; Obj L The objective function for predicting the right leaf node; Obj (L+R) The objective function for predicting the parent leaf node.
[0033] In the aforementioned wind turbine group runaway warning method in the same offshore wind farm, during the XGBoost model training process, the decision tree starts to grow from the root, splits the node into two new nodes on the left and right through the node condition, and distributes the parent node weight to the new node according to the node condition, and continues to split until it stops;
[0034] In order to prevent the tree generated by the XGBoost model from being too complex, resulting in a huge model and overfitting of the model, the tree stops growing when the Gain value is less than 0 and cannot bring gain to the objective function, or when the depth of the tree growth reaches the preset maximum depth, or when the number of leaf node samples is too small;
[0035] At the same time, the maximum depth, learning rate and weak estimator of the model are tuned, and the model hyperparameter adjustment is optimized using grid search to obtain the XGBoost model with the best prediction effect.
[0036] In the aforementioned method for early warning of group runaway of wind turbines in the same offshore wind farm, step 4: the data filling uses a generator through a GAIN network to estimate the missing data, uses a discriminator to discriminate the error between the estimated value and the true value, and continuously updates the parameters of the generator and the discriminator until the Nash equilibrium point is found, so that the generator and the discriminator loss functions obtain the best results.
[0037] In the above-mentioned method for early warning of wind turbine group overrun in the same offshore wind farm, the generator uses data The mask matrix M and noise Z are used as input. The mask matrix M consists of 0 or 1, which indicates the location of missing values in the data set. The model returns a Data after filling missing values The data X and the data after missing values The calculation formula is as follows:
[0038]
[0039]
[0040] In the formula, represents a random variable, and its value is specified as follows:
[0041]
[0042] Where M is a random variable with values of {0, 1}, X is a data vector, and * indicates that when M i When it is 0, it indicates an unobserved value.
[0043] In the above-mentioned method for early warning of wind turbine group flying in the same offshore wind farm, the discriminator uses the generator as an opponent for training the discriminator; the confrontation process between the generator and the discriminator is carried out using a random variable H; the element h in the random variable H depends on the distribution HM i =m, each input sample Input to the discriminator, based on the condition and H = m, discriminator The p parts in reflect the probability that it is real data. By defining different random variables H, different prompt information is given to the discriminator, so as to train generators with different distributions and select the optimal generator and discriminator.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] The present invention constructs an XGBoost model to accurately predict the average generator speed of a certain wind turbine unit, and completes the overspeed warning of the offshore wind turbine unit by monitoring its prediction difference and the threshold of the prediction value; the present invention also uses the GAIN network to fill in the lack of complete data sets, generates virtual data to fill the data set and verifies the overspeed warning model, thereby realizing overspeed monitoring of wind turbines that lack historical data, has a high monitoring accuracy, can reduce the hidden dangers of overspeed monitoring and warning costs of offshore generator sets, and is of great significance for improving the power generation efficiency of offshore wind turbines and improving the benefits of the units. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 This is a flow chart of the offshore wind turbine runaway warning of the present invention;
[0047] Figure 2 It is a correlation analysis diagram of the present invention;
[0048] Figure 3 It is a ranking diagram of the importance of characteristic parameters of the present invention;
[0049] Figure 4 It is a comparison chart of the predicted value and the actual value of the 3#6.8MW unit runaway monitoring with a comprehensive data set of the present invention;
[0050] Figure 5This is a comparison chart of the predicted value and actual value of the 61#8.3MW unit runaway monitoring with a comprehensive data set of the present invention. DETAILED DESCRIPTION
[0051] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.
[0052] Embodiment: A method for early warning of wind turbine group runaway in the same offshore wind farm, such as Figure 1 As shown, the specific steps are as follows:
[0053] Step 1: Obtain the operating data of offshore wind turbines, and perform data preprocessing, correlation analysis, and feature screening to obtain a high-quality, highly relevant data set;
[0054] In this embodiment, the offshore wind turbine operating data is collected by a SCADA system on the offshore wind turbine.
[0055] The data preprocessing is specifically as follows: taking the 3#6.8MW unit as the experimental object, first of all, binary features and categorical features should be excluded, such as the average unit yaw mode, the average unit operation mode, and the maximum value features, such as the maximum hydraulic system oil temperature. These data features have small changes and the amount of information contained in the data is small, so they need to be excluded first. At the same time, although sensors are arranged for some characteristic parameters, the corresponding data has not been collected, and these data features should all be excluded. Delete the data with an average generator speed below 60 revolutions per minute. Data under abnormal operating conditions will greatly affect the training effect of the model, and will also cause the evaluation index to be abnormally large in value, which cannot normally reflect the effect of model training. At the same time, there is no risk of running away when the wind turbine is not running.
[0056] like Figure 2 As shown in the figure, the correlation analysis specifically uses the Pearson correlation coefficient method to calculate the correlation coefficient between the feature parameters for preliminary screening; the feature screening uses the "gain" weight form in the XGBoost model to screen the remaining feature parameters according to the feature importance index, such as Figure 3 shown.
[0057] Step 2: Build an XGBoost model and input the data set into the XGBoost model for training until the model prediction effect reaches the best and the optimal XGBoost model is obtained;
[0058] Q=(x i ,y i )(|Q|=n,x i ∈R m ,y i ∈R);
[0059] The predicted value function in the prediction process is as follows:
[0060]
[0061] Where i represents the i-th sample in the training set; is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree;
[0062] The objective function is defined as follows:
[0063]
[0064] Where Obj is the objective function; i is the i-th sample in the training set; y i is the true value of the i-th sample; is the i-th sample x i The predicted value of is the loss function; Ω(f k ) is the regularization term;
[0065] The loss function is defined as follows:
[0066]
[0067] The regularization term is defined as follows:
[0068]
[0069] Where T is the number of leaf nodes in the decision tree; ω represents the score on each leaf node; γ and λ represent regularization parameters;
[0070] The XGBoos model function obtained by the t-th training is defined as follows:
[0071]
[0072] Among them, Obj (t) is the objective function obtained from the tth learning; Ω(f t ) is the regularization term of the XGBoos model function during the t-th learning; y i is the true value of the i-th sample; For the tth learning, the i-th sample x i The predicted value of; C is a constant;
[0073] The gain function formula is defined as follows:
[0074] Gain=Obj (L+R) -(Obj L +Obj L );
[0075] Where, Gain is the gain function; Obj L The objective function for predicting the left leaf node; Obj L The objective function for predicting the right leaf node; Obj (L+R) The objective function for predicting the parent leaf node.
[0076] During the model training process, the decision tree starts to grow from the root, splits the node into two new nodes on the left and right through the node conditions, and distributes the parent node weight to the new node according to the node conditions, and continues to split until it stops. In order to prevent the tree generated by the XGBoost model from being too complex, resulting in a huge model and overfitting of the model, when the Gain value is less than 0 and cannot bring gain to the objective function; when the depth of the tree growth reaches the preset maximum depth; or when the number of leaf node samples is too small, the tree stops growing and the model training is completed.
[0077] Step 3: using the optimal XGBoost model, predict the average generator speed of the offshore wind turbines, and monitor the difference between the predicted value and the true value to complete the runaway warning of the offshore wind turbines in the group;
[0078] After data preprocessing, data filling and feature screening, the data in the dataset is optimized for the three hyperparameters of the model, namely, the maximum depth (max_depth), learning rate (learning_rate) and weak estimators (n_estimators). The model hyperparameter adjustment is optimized using grid search (GridSearchCV) to obtain the XGBoost model with the best prediction effect and predict the average generator speed. The prediction effect of the model is then evaluated by the root mean square error (RMSE) and the mean absolute percentage error (MAPE).
[0079] Most of the offshore wind turbines in this data set can only provide characteristic parameters such as average wind direction, average cabin temperature, average outdoor temperature, average gearbox oil temperature, and average gearbox main bearing temperature. At the same time, the operating data of offshore wind turbines has a cycle of up to one year, which has a great relationship with time, and one month is used as a time parameter to participate in model training. Most of these characteristic parameters have a low correlation with the average generator speed, and it is difficult to accurately estimate the current generator speed. The accuracy of predicting the average generator speed is shown in Table 1.
[0080] Table 1
[0081] Unit Model Rated Power MAPE RMSE Unit Model Rated Power MAPE RMSE 1# 6.8MW 0.0315 12.681 4# 8.3MW 0.0353 12.700 18# 6.8MW 0.0498 17.808 12# 8.3MW 0.0388 16.195 29# 6.8MW 0.0399 15.337 13# 8.3MW 0.0432 13.760 38# 6.8MW 0.0378 12.906 27# 8.3MW 0.0456 14.532 49# 6.8MW 0.0453 17.786 44# 8.3MW 0.0395 13.258 51# 6.8MW 0.0474 16.295 50# 8.3MW 0.0467 14.598
[0082] As can be seen from the table above, the prediction results for the average generator speed using the original data set are low, and the runaway warning has high requirements for the correlation of feature parameters. The accuracy of the runaway monitoring model is about 95%, which has certain practical value, but in the case of high-speed operation, there are also problems such as failure to issue runaway detection warnings in time or frequent false alarms. After transfer learning, the accuracy will be lower, which is difficult to meet the actual wind turbine runaway warning needs.
[0083] After the data set is filled, the data features with a higher correlation with the average generator speed can be used for runaway warning. The runaway warning model is made based on the 3#6.8MW and 61#8.3MW wind turbines with complete data sets. Figure 4 and Figure 5 shown.
[0084] In order to prevent XGBoost from finding patterns and overfitting the algorithm based on time when training the model, the data is randomly shuffled. Figure 4 and Figure 5 It can be seen that the curves of the predicted value and the observed value of the average generator speed have a high degree of overlap, indicating that the selected characteristic parameters are very important to the model prediction results. In addition, the prediction effect of the flying car monitoring model is evaluated by using evaluation indicators. The RMSE of the 3#6.8MW unit is 1.1764, the MAPE is 0.0023, and the RMSE of the 61#8.3MW unit is 0.9085, and the MAPE is 0.0021.
[0085] The flying car monitoring model can predict the average generator speed with high accuracy, but this method has high requirements on the quality and quantity of the characteristic parameters input to the model. Using characteristic parameters with high correlation, the prediction accuracy of the flying car warning model is more than 98%, and the RMSE index is also greatly reduced. The evaluation indicators are all in a low range, indicating that the model has a very good prediction effect on the generator speed and has high practical value.
[0086] Step 4: Use the GAIN network to fill in the data of other offshore wind turbines with missing data, and then use transfer learning to migrate the trained optimal XGBoost model to other offshore wind turbines. Based on the operating data of other offshore wind turbines, the average generator speed is predicted, and then a runaway warning is issued.
[0087] Only two offshore wind turbines have comprehensive data sets. Due to the large amount of data and the unwillingness of companies to disclose data information, the data obtained by most units are not comprehensive. At the same time, the existing data sets themselves have some missing values every once in a while. Therefore, the quality of the data set is difficult to meet the needs of training the flying car fault warning model and testing transfer learning.
[0088] GAIN (Generative Adversarial Imputation Nets) is a network based on the GAN network framework that uses estimation to fill in missing values. Its basic principle is similar to that of the standard GAN. The generator is used to estimate the missing data, and the discriminator determines the error between the estimated value and the true value. The parameters of the generator and the discriminator are continuously updated until the Nash equilibrium point is found, so that the loss function of the generator and the discriminator obtains the best result.
[0089] The generator is used to observe each part of the real data, and then fill in the missing data according to the observation results, and output a complete vector after filling; the discriminator: accepts the complete vector and determines which part of the data is the real data and which part of the data is the generator; the GAIN network reveals the information of the missing part in the original data, allowing the discriminator to pay more attention to the prompted part, and then forces the generator to generate data closer to the real part to fill it.
[0090] The generator takes data The mask matrix M and noise Z are used as input. The mask matrix M consists of 0 or 1, which indicates the location of missing values in the data set. The model returns a Data after filling missing values The data And the data after missing values The calculation formula is as follows:
[0091]
[0092]
[0093] In the formula, represents a random variable, and its value is specified as follows:
[0094]
[0095] Where M i is a random variable with values {0, 1}, X i is a data vector, * indicates that when M i When it is 0, it indicates an unobserved value.
[0096] The discriminator uses the generator as an opponent to train the discriminator. Unlike the standard GAN network, the results generated by the GAN generator are either all real or all fake. The results given by the GAIN network generator are partly real and partly generated. The discriminator does not judge whether the entire vector is real or fake, but judges which part of the vector is real and which part is generated by predicting the m value in M.
[0097] The adversarial process between the generator and the discriminator is carried out using a random variable H; the element h in the random variable H depends on the distribution HM i =m, each input sample Input to the discriminator, based on the condition and H = m, discriminator The p parts in reflect the probability that it is real data. By defining different random variables H, different prompt information is given to the discriminator, so as to train generators with different distributions and select the optimal generator and discriminator.
[0098] The GAIN network was first applied to a small amount of missing data, showing a good filling effect. At the same time, since there are usually many units of the same model and brand in a single wind farm, their structures and principles are highly similar, and they operate in similar environments, resulting in a high degree of similarity in the data distribution and data features generated during their operation. Therefore, this paper fills in other data features based on the existing 3#6.8MW unit and 61#8.3MW unit.
[0099] The GAIN network has a high accuracy rate in filling large areas of data. The accuracy rate is tested using the existing unit parameters. The data is filled using units of the same model. The filling effect is shown in Table 2.
[0100] Table 2
[0101]
[0102]
[0103] First, use GAIN to fill the data set. The filled data set conforms to the average value and distribution of the original data and can be used as the corresponding training data set. The filled data set is used as the experimental object of transfer learning. And compared with the evaluation indicators of other models and the original data set, the practicality of transfer learning for the flying warning model is evaluated. The transfer learning model will be fine-tuned on the limited data of each unit to adapt to the new environment, so as to obtain better prediction performance when the new wind farm data is insufficient. The 1000 data after data filling in the offshore wind turbine lacking comprehensive historical data are used as the training set, and the remaining data are used as the test set. First, use the data preprocessing method of the offshore wind turbine flying warning model trained with comprehensive data sets 3#6.8MW and 61#8.3M in step 3 to preprocess the operating data of other wind turbines in the same wind farm. Then analyze the operating data of the wind turbine. The data of 1,000 wind turbines are input into a trained offshore wind turbine runaway warning model with a comprehensive data set. The model is used to fine-tune the data, and then the average generator speed of other wind turbines is predicted, thereby providing runaway warnings for other wind turbines that lack comprehensive data.
[0104] The results are shown in Table 3.
[0105] Table 3
[0106] Unit Model Rated Power MAPE RMSE Unit Model Rated Power MAPE RMSE 1# 6.8MW 0.0153 6.9280 4# 8.3MW 0.0109 4.0637 18# 6.8MW 0.0285 11.510 12# 8.3MW 0.0194 7.0061 29# 6.8MW 0.0118 4.2639 13# 8.3MW 0.0272 8.4166 38# 6.8MW 0.0167 5.4310 27# 8.3MW 0.0393 7.5234 49# 6.8MW 0.0207 7.5780 44# 8.3MW 0.0238 7.8258 51# 6.8MW 0.0098 3.3386 50# 8.3MW 0.0205 5.8472
[0107] The transfer learning effects of wind turbines of different models and locations are different, but compared with the original data set, the evaluation indicators RMSE and MAPE after the model transfer have been greatly improved due to the use of feature parameters with higher correlation coefficients. After transfer learning, the prediction accuracy of the average generator speed can reach more than 96%, and the full cycle data can be predicted after a small amount of data adjustment, meeting the actual operation requirements of the flying car warning model.
[0108] In summary, the present invention can improve the accuracy of runaway warning monitoring and reduce the hidden dangers of runaway monitoring and warning costs of offshore wind turbines.
Claims
1. A method for early warning of wind turbine group runaway in the same offshore wind farm, characterized in that: The following steps are involved: Step 1: Obtain the operating data of a group of offshore wind turbines, and perform data preprocessing, correlation analysis and feature screening on the operating data to obtain a data set; Step 2: Build an XGBoost model and input the data set into the XGBoost model for training until the model prediction effect reaches the best and the optimal XGBoost model is obtained; Step 3: using the optimal XGBoost model, predict the average generator speed of the offshore wind turbines, and monitor the difference between the predicted value and the true value to complete the runaway warning of the offshore wind turbines in the group; Step 4: Use the GAIN network to fill in the data of other offshore wind turbines with missing data, and then use transfer learning to migrate the trained optimal XGBoost model to other offshore wind turbines in the same wind farm, so as to achieve runaway warning for all offshore wind turbines in the entire wind farm.
2. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 1, characterized in that: In step 1, the offshore wind turbine operating data is collected by a SCADA system on the offshore wind turbine.
3. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 1, characterized in that: The data preprocessing includes eliminating the maximum value features and category features in the offshore wind turbine operating data, removing feature parameters with a large proportion of zero values and missing values, and deleting data in abnormal working conditions of the wind turbine.
4. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 1, characterized in that: The correlation analysis specifically uses the Pearson correlation coefficient method to calculate the correlation coefficients between feature parameters for preliminary screening; the feature screening is to screen the remaining feature parameters according to the feature importance index.
5. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 1, characterized in that: In step 2, the process of building the XGBoost model is: Given a model with n samples and m features: Q=(x i ,y i )(|Q|=n,x i ∈R m ,y i ∈R); The predicted value function in the prediction process is as follows: Among them, i represents the i-th sample in the training set; is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree; The objective function is defined as follows: Where Obj is the objective function; i is the i-th sample in the training set; y i is the true value of the i-th sample; is the i-th sample x i The predicted value of is the loss function; Ω(f k ) is the regularization term; The loss function is defined as follows: The regularization term is defined as follows: Where T is the number of leaf nodes in the decision tree; ω represents the score on each leaf node; γ and λ represent regularization parameters; The XGBoos model function obtained by the t-th training is defined as follows: Among them, Obj (t) is the objective function obtained from the tth learning; Ω(f t ) is the regularization term of the XGBoos model function during the t-th learning; y i is the true value of the i-th sample; For the tth learning, the i-th sample x i The predicted value of; C is a constant; The gain function formula is defined as follows: Gain=Obj (L+R) -(Obj L +Obj L ); Where, Gain is the gain function; Obj L The objective function for predicting the left leaf node; Obj L The objective function for predicting the right leaf node; Obj (L+R) The objective function for predicting the parent leaf node.
6. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 5, characterized in that: During the XGBoost model training process, the decision tree grows from the root, splits the node into two new nodes on the left and right according to the node condition, and distributes the parent node weight to the new node according to the node condition, and continues to split until it stops; In order to prevent the tree generated by the XGBoost model from being too complex, resulting in a huge model and overfitting of the model, the tree stops growing when the Gain value is less than 0 and cannot bring gain to the objective function, or when the depth of the tree growth reaches the preset maximum depth, or when the number of leaf node samples is too small; At the same time, the maximum depth, learning rate and weak estimator of the model are tuned, and the model hyperparameter adjustment is optimized using grid search to obtain the XGBoost model with the best prediction effect.
7. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 1, characterized in that: Step 4: The data filling uses the generator through the GAIN network to estimate the missing data, uses the discriminator to determine the error between the estimated value and the true value, and continuously updates the parameters of the generator and the discriminator until the Nash equilibrium point is found, so that the generator and discriminator loss functions obtain the best results.
8. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 7, characterized in that: The generator takes data The mask matrix M and noise Z are used as input. The mask matrix M consists of 0 or 1, which indicates the location of missing values in the data set. The model returns a Data after filling missing values The data And the data after missing values The calculation formula is as follows: In the formula, represents a random variable, and its value is specified as follows: Where M i is a random variable with values {0, 1}, X i is a data vector, * indicates that when M i When it is 0, it indicates an unobserved value.
9. The method for early warning of wind turbine group runaway in the same offshore wind farm according to claim 8, characterized in that: The discriminator uses the generator as an opponent to train the discriminator; the adversarial process between the generator and the discriminator is carried out using a random variable H; the element h in the random variable H depends on the distribution HM i =m, each input sample Input to the discriminator, based on the condition and H = m, discriminator The p parts in reflect the probability that it is real data. By defining different random variables H, different prompt information is given to the discriminator, so as to train generators with different distributions and select the optimal generator and discriminator.