Fan gear box oil temperature prediction method and system based on PCA, MIC and fusion model

Abstract

This invention relates to the field of wind turbine gearbox oil temperature prediction, specifically to a method and system for predicting wind turbine gearbox oil temperature based on PCA, MIC, and a fusion model. The technical solution includes: filtering initial features using MIC and a correlation matrix; performing spatial transformation on the initial features using PCA to obtain principal component vectors with information content greater than a set threshold; combining the relevant principal component vectors with the filtered initial features to obtain a first feature set; performing secondary correlation analysis on the first feature set, and filtering out variables in the first feature set whose correlation is greater than a set threshold using MIC; adding the first two principal component vectors to the first feature set to obtain a second feature set; training and cross-validating the first, second, and third models using the second feature set, fusing the weights to obtain a fusion model; and predicting the wind turbine gearbox oil temperature using the fusion model. This invention is applicable to gearbox oil temperature prediction.

Description

Technical Field

[0001] This invention relates to the field of wind turbine gearbox oil temperature prediction, specifically to a method and system for predicting wind turbine gearbox oil temperature based on PCA (Principal Component Analysis) and MIC (Maximal Information Coefficient) and a fusion model. Background Technology

[0002] In the actual operation and maintenance of wind turbine units, both external conditions such as weather, collisions and friction, and structural changes in components over long-term operation, and internal conditions such as electrostatic heat during component operation and internal circuit failures causing shutdowns, all affect the stable production of wind turbine units. Therefore, in order to provide a more stable guarantee for safe power production and to conduct a more comprehensive assessment of unit performance and heat dissipation, monitoring the temperature of each component of the generator unit to assist in fault diagnosis has become a common practice.

[0003] As a core component for energy transfer in a generator set, gearbox condition monitoring is crucial for ensuring the safety and stability of the generator's operation. Gearbox oil temperature directly reflects the gearbox's condition, making it a key method for ensuring stable generator operation during maintenance. With the increasing application of machine learning and algorithms in sequence prediction, predicting temperature in advance using machine learning algorithms for early warning has become a popular research topic, aiming to provide timely information and assessments of generator status.

[0004] Common early warning methods include collecting additional relevant data such as the operating conditions of the surrounding environment, power, voltage, and current. This data is used as supplementary auxiliary features, and then commonly used neural networks (BP, CNN, RNN, etc.) are trained on the oil temperature data. Alternatively, feature recognition methods can be employed to predict oil temperature, including LightGBM, LSTM, and dynamic neural network methods, as well as methods using residual distribution as a loss function.

[0005] However, due to the uncertainty of traditional data, and the noise and insufficient information between data points, actual model training is often interfered with. Traditional neural networks suffer from the black-box nature and severe overfitting in identifying regular data, resulting in high time and effort consumption. Furthermore, using a single model for prediction also leads to low accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting wind turbine gearbox oil temperature based on PCA, MIC and a fusion model, which greatly improves the accuracy of wind turbine gearbox oil temperature prediction.

[0007] This invention achieves the above objective by adopting the following technical solution: a method for predicting wind turbine gearbox oil temperature based on PCA, MIC, and a fusion model, comprising:

[0008] Initial data feature selection and reconstruction:

[0009] The initial features are filtered using the MIC and correlation matrix to obtain the filtered initial features;

[0010] PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold.

[0011] The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set;

[0012] Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation.

[0013] The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set.

[0014] Model building and prediction:

[0015] The first, second, and third models are trained and cross-validated using the second feature set. After cross-validation, the different models are multiplied by their corresponding weights and then summed to obtain the fusion model.

[0016] The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

[0017] Furthermore, the method also includes:

[0018] When determining the weights, the inverse ratio of the error of the corresponding model on the validation set is used as the determination criterion for the weights. Therefore, the formula for determining the weights is:

[0019]

[0020] The error and weights of the two models, e l w l These represent the error and weights of the third model, respectively.

[0021] Furthermore, the first model is an extreme stochastic boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model.

[0022] Furthermore, the method also includes:

[0023] During model training, the Adam algorithm and stochastic gradient descent algorithm are used to optimize the hyperparameters.

[0024] During cross-validation, cross-validation is performed on each model, and the mean of the error and the mean of the result are taken as the input values ​​of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

[0025] A wind turbine gearbox oil temperature prediction system based on PCA, MIC, and a fusion model is provided to implement the aforementioned wind turbine gearbox oil temperature prediction method based on PCA, MIC, and a fusion model. The system includes:

[0026] The initial data feature filtering and reconstruction module is used to filter the initial features through MIC and correlation matrix to obtain the filtered initial features;

[0027] PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold.

[0028] The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set;

[0029] Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation.

[0030] The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set.

[0031] The model building and prediction module is used to train and cross-validate the first, second, and third models using the second feature set. After cross-validation, the different models are multiplied by their corresponding weights and then added together to obtain the fusion model.

[0032] The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

[0033] Furthermore, the model building and prediction module is also used to determine the weights by taking the inverse ratio of the corresponding model's error on the validation set as the weight confirmation criterion. Therefore, the weight confirmation formula is:

[0034] Where e x w x These represent the error and weights of the first model, e. r w rThese represent the error and weights of the second model, respectively. l w l These represent the error and weights of the third model, respectively.

[0035] Furthermore, the first model is an extreme stochastic boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model.

[0036] Furthermore, the model building and prediction module is also used to optimize hyperparameters using the Adam algorithm and stochastic gradient descent algorithm during model training.

[0037] During cross-validation, cross-validation is performed on each model, and the mean of the error and the mean of the result are taken as the input values ​​of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

[0038] The beneficial effects of this invention are as follows:

[0039] In the feature processing stage, this invention uses PCA and MIC to screen and reconstruct features. For the obtained first feature set, a secondary correlation analysis is performed and variables with large correlations are screened out by MIC to eliminate redundant noise and purify the features. Furthermore, the first two principal component vectors are added to the first feature set to obtain the second feature set, which facilitates the learning of highly condensed variable information during model training.

[0040] In the process of model building, this invention integrates multiple basic models. Different weights are assigned to different basic models according to their prediction effects. Models with better performance are given higher weights, and the smaller the error, the higher the weight value. The fusion model is formed by multiplying multiple basic models by their weights and then adding them together. Therefore, using the fusion model to predict the oil temperature of the wind turbine gearbox can greatly improve the accuracy of the prediction. Attached Figure Description

[0041] Figure 1 A flowchart illustrating the feature selection and reconstruction process provided in this embodiment of the invention;

[0042] Figure 2 A flowchart for model establishment provided in the embodiments of the present invention;

[0043] Figure 3 The values ​​of the pc1-pc5 feature vectors provided in the embodiments of the present invention;

[0044] Figure 4 This is a comparison chart of the predicted results and actual results provided in the embodiments of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0046] This invention provides a method for predicting wind turbine gearbox oil temperature based on PCA, MIC, and a fusion model, including:

[0047] Initial data feature selection and reconstruction:

[0048] The initial features are filtered using the MIC and correlation matrix to obtain the filtered initial features;

[0049] PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold.

[0050] The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set;

[0051] Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation.

[0052] The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set.

[0053] Model building and prediction:

[0054] The first, second, and third models are trained and cross-validated using the second feature set. After cross-validation, the different models are multiplied by their corresponding weights and then summed to obtain the fusion model.

[0055] The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

[0056] In one embodiment of the present invention, the initial data features are screened and reconstructed as follows: Figure 1As shown, in the feature processing stage, the initial features are divided into two parts: first, using PCA to perform spatial transformation to obtain principal component vectors with high information content (e.g., obtaining pc1-pc10); and second, using MIC and correlation matrix methods to filter the initial features (e.g., obtaining the top 20 MIC values ​​with the target variable). The top-ranked principal component vectors are then combined with the filtered initial features to obtain the first feature set. Subsequently, for the first feature set, secondary correlation analysis is performed, and variables with high correlations are filtered out using MIC to eliminate redundant noise and purify the feature set. Since the first two pc variables (e.g., pc1 and pc2) contain a large amount of principal component information, they are retained to assist the model in generalizing the learning of highly condensed variable information.

[0057] In one embodiment of the present invention, model building and prediction are as follows: Figure 2 As shown, to improve the overall predictive performance and increase the robustness of the model, we fused several models of different types. The basic models used were Extreme Gradient Boosting Regressor (XGB), Random Forest Regressor, and LSTM (Long Short-Term Memory) neural network model.

[0058] For prediction problems involving continuous data, a common approach is the weighted average fusion method. This method ranks different base models according to their prediction performance, assigning higher weights to models with better performance. The smaller the error, the higher the weight. The weights of all models are summed to 1. The XRL fusion model is formed by multiplying the base models by their weights and then summing the results. When determining the weights, the inverse ratio of their errors on the validation set is used as the weight confirmation criterion. Therefore, the weight confirmation formula is:

[0059] Where e x w x These represent the error and weights of the first model, e. r w r These represent the error and weights of the second model, respectively. l w l These represent the error and weights of the third model, respectively.

[0060] The first model is an extreme random boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model. Alternatively, several algorithms such as K-nearest neighbors, bagging, LGBM, decision trees, gradient boosting, and extreme random trees can be used as the base model and fused together.

[0061] In one embodiment of the present invention, the method further includes:

[0062] During model training, the Adam algorithm and stochastic gradient descent algorithm are used to optimize the hyperparameters.

[0063] During cross-validation, cross-validation is performed on each model, and the mean of the error and the mean of the result are taken as the input values ​​of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

[0064] This invention also provides a wind turbine gearbox oil temperature prediction system based on PCA, MIC, and a fusion model, used to implement the aforementioned wind turbine gearbox oil temperature prediction method based on PCA, MIC, and a fusion model. The system includes:

[0065] The initial data feature filtering and reconstruction module is used to filter the initial features through MIC and correlation matrix to obtain the filtered initial features;

[0066] PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold.

[0067] The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set;

[0068] Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation.

[0069] The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set.

[0070] The model building and prediction module is used to train and cross-validate the first, second, and third models using the second feature set. After cross-validation, the different models are multiplied by their corresponding weights and then added together to obtain the fusion model.

[0071] The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

[0072] In one embodiment of the present invention, the model building and prediction module is further configured to, when confirming weights, take the inverse ratio of the error of the corresponding model on the validation set as the confirmation criterion for the weights, and the weight confirmation formula is:

[0073] Where e x w x These represent the error and weights of the first model, e. r w rThese represent the error and weights of the second model, respectively. l w l These represent the error and weights of the third model, respectively.

[0074] In one embodiment of the present invention, the first model is an extreme stochastic boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model.

[0075] In one embodiment of the present invention, the model building and prediction module is further used to optimize the hyperparameters using the Adam algorithm and the stochastic gradient descent algorithm during the model training process.

[0076] During cross-validation, cross-validation is performed on each model, and the mean of the error and the mean of the result are taken as the input values ​​of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

[0077] The present invention will now be described in further detail with reference to specific embodiments.

[0078] Taking the gearbox oil temperature data of a wind turbine generator set at a power station in Southwest China as an example, the data varies with the power of the wind turbine generator set over time. The data includes 38 initial features such as system time, temperature of various grid-side components, motor speed, and ambient temperature, and the data volume is more than 40,000 data points for a whole year, with each time node being ten minutes.

[0079] The initial features are shown in the table below:

[0080]

[0081]

[0082] (I) Feature Recombination and Screening

[0083] Since the original data has 38 features, the spatially transformed data obtained using the PCA method has 38 rows and 38 columns. Each row corresponds to the vector value of the corresponding feature mapped on the spatially transformed coordinate axis, and each column corresponds to the mapping value of the original feature on different PC coordinate axes. Figure 3 The values ​​are pc1 to pc5. Negative values ​​indicate that the corresponding initial feature contributes a negative direction vector to this pc feature, while positive values ​​contribute a positive vector. A value of 0 indicates that the feature is perpendicular to this feature vector space and is therefore irrelevant to the construction of this feature vector.

[0084] When filtering data features, in addition to the MIC (Maximum Information Count) indicator, we also considered the content of the correlation matrix. From the correlation matrix, we can identify features that contribute to the prediction target and features that are noisy to each other. The table below shows the correlation matrix between each pc vector value and the prediction target. The values ​​in the matrix range from [-1, 1], and the closer the absolute value is to 1, the higher the correlation.

[0085]

[0086]

[0087] The table below shows the MIC index values ​​for each pc feature, with a range of [0,1]. The larger the value, the more information the feature contains.

[0088] pc1 pc2 pc3 pc4 PC5 pc6 pc7 pc8 pc9 pc10 pc11 pc12 pc13 0.96 0.40 0.46 0.45 0.21 0.30 0.37 0.18 0.24 0.09 0.18 0.11 0.16 pc14 pc15 pc16 pc17 pc18 pc19 pc20 pc21 pc22 pc23 pc24 pc25 pc26 0.09 0.12 0.20 0.17 0.13 0.19 0.14 0.13 0.12 0.14 0.12 0.18 0.10 pc27 pc28 pc29 pc30 pc31 pc32 pc33 pc34 pc35 pc36 pc37 pc38 … 0.12 0.09 0.11 0.09 0.08 0.15 0.10 0.08 0.10 0.09 0.10 0.13 …

[0089] (II) Algorithm Model Training

[0090] 1. First, the dataset after feature processing (i.e., the second dataset in this embodiment) is randomly divided into training and test sets at a ratio of 4:1 (80% of the data is the training set and 20% of the data is the test set).

[0091] 2. Subsequently, the three basic models (extreme gradient boosting model, random forest model, and long short-term memory neural network model) were trained and learned on the entire training dataset. During training, the Adam and stochastic gradient descent (SGD) methods were used to optimize the hyperparameters. The update rule for the stochastic gradient boosting optimization algorithm is as follows:

[0092]

[0093] 3. During training, the batch size controls the training speed, but may affect the training results of the model. It is a key factor in determining the training time and computational resource consumption. We will try to adopt a larger batch size when we reasonably consider the current resources and time.

[0094] 4. The final prediction result is influenced by the prediction results and errors of each of its constituent base models. After the optimizer adjusts the parameters, cross-validation is performed on each base model, and the mean of the errors and the mean of the results from the validation sets are used as the input values ​​for the final result. The higher the error, the lower the weight of its prediction in the final result. When determining the weights, the inverse ratio of the error on the validation set is used as the weight determination criterion. Therefore, the weight determination formula is:

[0095]

[0096] 5. Multiply the outputs of each model by their respective weights and then fuse them together to obtain the final output:

[0097] OUtput = w x *Out x +w r *Out r +w l *Out l

[0098] Where w x This indicates that extreme randomness increases the weights corresponding to the regression model, Out x w represents the output of an extreme random boost regression model. r Represents the weights corresponding to the random forest regression model, Out r w represents the output of the random forest regression model. l Out represents the weights corresponding to the Long Short-Term Memory (LSTM) neural network model. l This represents the output of the Long Short-Term Memory (LSTM) neural network model. The final predicted temperature and the actual temperature are as follows: Figure 4 As shown.

[0099] The prediction performance under different algorithms is compared in the table below.

[0100] Model R-squared (goodness of fit) RMSE (Root Mean Square Error) Adjusted R-Squared PCA_MIC_XRL 0.95 1.93 0.95 PCA_MIC_XRL+pc6,pc7 0.95 1.97 0.95 Random Forest 0.93 2.37 0.93 XGB Returns 0.93 2.79 0.93 LSTM Neural Network 0.90 6.33 0.90

[0101] As shown in the table above, the fusion model PCA_MIC_XRL or the fusion model with added feature vectors pc6 and pc7 has better fit, root mean square error, and adjusted fit than other single models.

[0102] In summary, this invention, in the feature processing stage, comprehensively considers the main components, information content, and relevance of features, thereby increasing the availability of variables; in the algorithm model building stage, it comprehensively considers the error of the basic model to determine the weights, thereby enhancing the stability of the model; it is easy to apply, highly reusable, and greatly improves the accuracy of wind turbine gearbox oil temperature prediction.

Claims

1. A method for predicting wind turbine gearbox oil temperature based on PCA, MIC, and a fusion model, characterized in that, include: Initial data feature selection and reconstruction: The initial features are filtered using the MIC and correlation matrix to obtain the filtered initial features; PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold. The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set; Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation. The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set. Model building and prediction: The first model, the second model, and the third model are trained and cross-validated using the second feature set. After cross-validation, the outputs of the different models are multiplied by their corresponding weights and then summed to obtain the output of the fusion model. The first model is an extreme stochastic boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model. The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

2. The method for predicting wind turbine gearbox oil temperature based on PCA, MIC, and their fusion model as described in claim 1, is characterized in that, The method also includes: When determining the weights, the inverse ratio of the error of the corresponding model on the validation set is used as the determination criterion for the weights. Therefore, the formula for determining the weights is: ,in , These are the error and weights of the first model, respectively. , These represent the error and weights of the second model, respectively. , These represent the error and weights of the third model, respectively.

3. The method for predicting wind turbine gearbox oil temperature based on PCA and MIC and their fusion model according to claim 1, characterized in that, The method also includes: During model training, the Adam algorithm and stochastic gradient descent algorithm are used to optimize the hyperparameters.

4. The method for predicting wind turbine gearbox oil temperature based on PCA, MIC, and their fusion model as described in claim 1, characterized in that, The method also includes: During cross-validation, cross-validation is performed on each model, and the mean of the error and the mean of the result are taken as the input values ​​of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

5. A wind turbine gearbox oil temperature prediction system based on PCA, MIC, and a fusion model, used to implement the wind turbine gearbox oil temperature prediction method based on PCA, MIC, and a fusion model as described in any one of claims 1-4, characterized in that, The system includes: The initial data feature filtering and reconstruction module is used to filter the initial features through MIC and correlation matrix to obtain the filtered initial features; PCA is used to spatially transform the initial features to obtain principal component vectors with more information than a set threshold. The principal component vectors with information content greater than the set threshold are combined with the filtered initial features to obtain the first feature set; Perform secondary correlation analysis on the first feature set, and use MIC to filter out variables in the first feature set whose correlation with each other is greater than the set threshold correlation. The principal component vectors are sorted from largest to smallest information content, and the first two principal component vectors are added to the first feature set to obtain the second feature set. The model building and prediction module is used to train and cross-validate the first model, the second model, and the third model using the second feature set. After cross-validation, the outputs of different models are multiplied by their corresponding weights and then summed to obtain the output of the fusion model. The first model is an extreme stochastic boosting regression model, the second model is a random forest regression model, and the third model is a long short-term memory neural network model. The oil temperature of the wind turbine gearbox is predicted by using a fusion model.

6. The wind turbine gearbox oil temperature prediction system based on PCA and MIC and their fusion model as described in claim 5, characterized in that, The model building and prediction module is also used to determine weights by taking the inverse ratio of the corresponding model's error on the validation set as the weight determination criterion. The weight determination formula is as follows: ,in , These are the error and weights of the first model, respectively. , These represent the error and weights of the second model, respectively. , These represent the error and weights of the third model, respectively.

7. The wind turbine gearbox oil temperature prediction system based on PCA and MIC and their fusion model according to claim 5, characterized in that, The model building and prediction module is also used to optimize hyperparameters during model training using the Adam algorithm and stochastic gradient descent algorithm.

8. The wind turbine gearbox oil temperature prediction system based on PCA and MIC and their fusion model according to claim 5, characterized in that, The model building and prediction module is also used to perform cross-validation on each model during the cross-validation process, and take the mean of the error and the mean of the result of the validation as the input value of the final result. The higher the error, the lower the proportion of its prediction result in the final result.

Images