CN117077532A - Multi-model fusion method for wind turbine life prediction - Google Patents

Abstract

The present invention provides a multi-model fusion method for wind turbine life prediction, which can comprehensively use multiple data and models to improve the stability and reliability of life prediction. It includes the following steps: S101, data collection and preprocessing S102, model construction and model training S103, model fusion and model evaluation S104, model application and model update.

Description

Multi-model fusion method for wind turbine life prediction

Technical field

The present invention relates to the technical field of wind turbine life, specifically a multi-model fusion method for wind turbine life prediction.

Background technique

Wind turbines are devices that convert wind energy into electrical energy. Its life prediction is an important part of wind farm operation and maintenance, which is related to the safety, reliability and economy of wind farms. The life of wind turbines is affected by many factors, such as structure, material, load, environment, etc. Therefore, life prediction is technically difficult and requires the comprehensive use of multiple methods and data.

At present, the main technologies for predicting the life of wind turbines include the following:

Life prediction technology based on cumulative fatigue damage theory: This technology uses Miner's cumulative fatigue damage theory to calculate the degree of damage and remaining life of key components of the wind turbine based on the load spectrum and material properties of the wind turbine. This technique requires the acquisition of accurate load data and material parameters, and has limited applicability to complex nonlinear damage mechanisms and multi-axial loading situations.

Life prediction technology based on condition monitoring and operation monitoring: This technology uses the fault warning and development trend prediction information given by the condition monitoring system and operation monitoring system, combined with traditional structural life assessment technology, to achieve a more comprehensive and accurate assessment of actual operating equipment. evaluation of remaining life. This technology requires the acquisition of a large amount of real-time data and historical data, and has high requirements for data quality and analysis methods.

Life prediction technology based on machine learning: This technology uses machine learning models, such as linear regression, support vector machines, neural networks, random forests, etc., to establish a life prediction model based on the historical data and current data of wind turbines, and predict the future life of the wind turbine. Make predictions. This technology requires obtaining sufficient training data and test data, and has higher requirements for model selection and parameter adjustment.

Each of the above technologies has its advantages and disadvantages, but none of them can make full use of the multi-source data and various characteristics of wind turbines, nor can they effectively deal with issues such as data incompleteness, uncertainty, and nonlinearity.

Contents of the invention

In response to the above problems, the present invention provides a multi-model fusion method for wind turbine life prediction, which can comprehensively use multiple data and models to improve the stability and reliability of life prediction.

The multi-model fusion method for wind turbine life prediction is characterized by including the following steps:

S101, data collection and preprocessing

Collect data from various data sources such as wind turbine operating status, environmental parameters, and fault records, and perform preprocessing operations such as data cleaning, data normalization, and feature extraction to obtain a data set suitable for model input;

S102, model construction and model training

According to different data characteristics and prediction goals, select an appropriate machine learning model, that is, model selection, and conduct model training on the data sets respectively to obtain several sub-models;

S103, model fusion and model evaluation

Use one or more model fusion methods to synthesize the output of multiple sub-models to obtain the final life prediction results, and evaluate the performance and accuracy of the model through cross-validation and error analysis methods;

S104, model application and model update

Deploy the model to the monitoring system of the wind farm to predict the life of the wind turbine regularly, and adjust and update the model based on actual operating conditions and feedback information.

It is further characterized by:

In the data collection step, various data are obtained from wind turbines in real time or regularly using sensors, monitoring systems, and log files, and stored in a database or cloud platform;

In the data cleaning step, anomaly detection, missing value filling, and noise filtering methods are used to perform quality checks and corrections on the data to remove or correct incomplete, inaccurate, or irrelevant data;

In the data normalization step, maximum and minimum value normalization, standardization, and logarithmic transformation methods are used to scale and distribute the data to eliminate the dimensional and skew effects of the data;

In the feature extraction step, principal component analysis, factor analysis, and wavelet transform methods are used to reduce the dimensionality and transform the data to extract the main features and information of the data;

In the model selection step, an appropriate machine learning model is selected based on the type and distribution of data, as well as the predicted tasks and indicators;

In the model training step, use the training set and validation set, as well as the corresponding loss function and optimization algorithm, to initialize and iteratively update the parameters of the model, and use regularization and early stopping methods to prevent the model from overfitting or underfitting;

In the model fusion step, select an appropriate model fusion method based on the type and output of the sub-models, and use the complementarity and diversity between sub-models to improve the stability and reliability of life prediction;

In the model evaluation step, the test set or new data, as well as the corresponding evaluation indicators and methods, are used to evaluate the prediction results of the model and test the performance and accuracy of the model;

In the model update step, the model is adjusted and updated based on the actual operating conditions and feedback information of the wind turbine.

After adopting the above technical solution, this method can comprehensively utilize multi-source data of wind turbines, such as operating status data, environmental parameter data, and fault record data, extract effective information and characteristics of the data, and increase the utilization rate and value of the data;

This method can comprehensively use a variety of machine learning models, such as linear regression, support vector machines, neural networks, and random forests, to build multiple sub-models, and use model fusion methods to take advantage of the complementarity and diversity between sub-models to improve lifespan. Forecast stability and reliability;

This method can predict the life of the wind turbine regularly or in real time based on the operational needs and planning of the wind farm, and display the prediction results on the monitoring interface or send them to relevant personnel or departments for subsequent analysis and processing;

This method can adjust and update the model based on the actual operating conditions and feedback information of the wind turbine, so that the model can maintain the latest and optimal status and improve the accuracy and robustness of the prediction.

Description of the drawings

Figure 1 is a flow chart of the method of the present invention;

Figure 2 is a schematic diagram of data collection and preprocessing of the present invention;

Figure 3 is a schematic diagram of model construction and model training of the present invention;

Figure 4 is a schematic diagram of model fusion and model evaluation of the present invention;

Figure 5 is a schematic diagram of model application and model updating in the present invention.

Detailed ways

The multi-model fusion method for wind turbine life prediction is shown in Figure 1, which includes the following steps:

S101: Data collection and preprocessing;

S102: Model construction and model training;

S103: Model fusion and model evaluation;

S104: Model application and model update.

The specific implementation is as follows:

S101: Data collection and preprocessing, as shown in Figure 2, this step includes the following sub-steps:

S1011: Data collection

Various data are obtained from wind turbines in real time or periodically using sensors, monitoring systems, log files, etc., and stored in databases or cloud platforms. For example, operating status data such as speed, power, temperature, vibration, and noise can be obtained from wind turbines, as well as environmental parameter data such as wind speed, wind direction, humidity, and air pressure, as well as fault record data such as fault type and fault time. These data can reflect the operating status and life characteristics of wind turbines, and can also be used as input for model training and prediction;

S1012: Data cleaning

Use methods such as anomaly detection, missing value filling, and noise filtering to perform quality checks and corrections on data to remove or correct incomplete, inaccurate, or irrelevant data. For example, you can use methods such as boxplots or the 3σ rule to detect and delete outliers or outliers; you can use methods such as mean interpolation or K nearest neighbor interpolation to fill in missing values; you can use methods such as sliding average or Kalman filtering to filter Remove noise. These methods can improve the quality and availability of data and avoid errors in model training and prediction;

S1013: Data normalization

Use methods such as maximum and minimum normalization, standardization, and logarithmic transformation to scale and distribute the data to eliminate the dimensional and skew effects of the data. For example, you can use the maximum and minimum value normalization method to convert the data into a value in the interval [0,1]; you can use the standardization method to convert the data into a normal distribution with a mean of 0 and a standard deviation of 1; you can use Logarithmic transformation method converts data into logarithmic form to reduce the degree of skewness. These methods can make the data meet the input requirements of the model and also improve the efficiency of model training and prediction;

S1014: Feature extraction

Use principal component analysis, factor analysis, wavelet transform and other methods to reduce the dimensionality and transform the data and extract the main features and information of the data. For example, you can use the principal component analysis method to convert high-dimensional features into low-dimensional features and retain the maximum amount of information; you can use the factor analysis method to convert multiple related features into a few irrelevant factors and reveal potential Causal relationship; wavelet transform method can be used to convert time domain features into frequency domain features and highlight local changes. These methods can reduce the redundancy and complexity of data, and also extract effective information from the data.

S102: Model construction and training, as shown in Figure 3. This step includes the following sub-steps:

S1021: Model selection

According to different data characteristics and prediction goals, select appropriate machine learning models, such as linear regression (LR), support vector machine (SVM), neural network (NN), random forest (RF), etc. For example, an appropriate machine learning model can be selected based on the type and distribution of data, as well as the prediction tasks and indicators, such as linear regression for fitting continuous value predictions of linear relationships, support vector machines for classification of nonlinear relationships, or Regression prediction, neural network is used to fit non-linear prediction of complex relationships, random forest is used to process ensemble learning prediction of high-dimensional features, etc. These models can be constructed and set up according to the characteristics and needs of the data;

S1022: Model training

Use the training set and validation set, as well as the corresponding loss function and optimization algorithm, to initialize parameters and iteratively update the model so that the model can fit the data and achieve optimal performance. At the same time, methods such as regularization and early stopping are used to prevent model overfitting or underfitting. For example, the following parameters and methods can be used:

Linear regression model: Use mean square error as the loss function, gradient descent method as the optimization algorithm, and L2 regularization to prevent overfitting. The parameters of this model include intercept terms and slope terms, which can be solved by the least squares method;

Support vector machine model: Use the 0-1 loss function as the loss function, use the sequential minimum optimization method as the optimization algorithm, and use cross-validation to select the kernel function and penalty parameters. The parameters of this model include the normal vector and intercept term of the classification hyperplane, as well as the type and parameters of the kernel function, which can be solved by maximizing the interval;

Neural network model: Use the cross-entropy loss function as the loss function, use the stochastic gradient descent method as the optimization algorithm, and use early stopping to prevent overfitting. The parameters of the model include the weights and biases of each layer, as well as the type and parameters of the activation function, which can be updated through the backpropagation algorithm;

Random forest model: uses Gini impurity as the loss function, adaptive boosting as the optimization algorithm, and cross-validation to select the number and depth of trees. The parameters of the model include the structure and splitting rules of each decision tree, as well as the proportion of training sets and feature sets generated by the bootstrap method and the random subspace method, which can be generated by the adaptive enhancement method.

S103: Model fusion and evaluation, as shown in Figure 4, this step includes the following sub-steps:

S1031: Model fusion

Use one or more model fusion methods, such as weighted average, voting method, stacking method, etc., to synthesize the output of multiple sub-models, and use the complementarity and diversity between sub-models to improve the stability and stability of life prediction Reliability; for example, the following methods can be used:

Model fusion: Using the stacking method, the outputs of the four sub-models are used as new features and input into a logistic regression model to obtain the final life prediction result. This method can use the logistic regression model to re-learn the sub-model output, thereby improving the accuracy and robustness of predictions;

S1032: Model evaluation

Use the test set or new data, as well as the corresponding evaluation indicators and methods, to evaluate the prediction results of the model and test the performance and accuracy of the model; for example, the following indicators and methods can be used:

Model evaluation: Use indicators such as mean square error, mean absolute error, and correlation coefficient to measure the error and correlation between predicted values and true values; use indicators such as confusion matrix, accuracy, and recall to measure the correctness and accuracy of classification predictions. Completeness etc. These indicators and methods can reflect the strengths and weaknesses of the model and the direction of improvement.

S104: Model application and update, as shown in Figure 5. This step includes the following sub-steps:

S1041: Model deployment

Convert the model into a format and interface suitable for the wind farm monitoring system and embed it into the wind farm monitoring system so that it can receive data and output prediction results; for example, the following frameworks and tools can be used:

Model deployment: Use frameworks such as TensorFlow Lite or ONNX to convert the model into a lightweight executable file or service and embed it into the wind farm monitoring system. These frameworks and tools can make models more compatible and portable;

S1042: Model prediction

According to the operational needs and planning of the wind farm, the life span of the wind turbine is predicted regularly or in real time, and the prediction results are displayed on the monitoring interface or sent to relevant personnel or departments for subsequent analysis and processing; for example, it can Use the following method:

Model prediction: Predict the life of the wind turbine according to the operating objectives and constraints of the wind farm, such as power generation, cost, safety, etc., and display the prediction results on the monitoring interface in the form of charts or reports, or send them to relevant personnel Or department, such as operation and maintenance personnel, management personnel, etc. These methods can enable wind farms to understand the life status of wind turbines in a timely manner and make corresponding decisions and operations;

S1043: Model update

The model is adjusted and updated based on the actual operating conditions and feedback information of the wind turbine; for example, the following methods can be used:

Model update: Adjust and update the model based on the actual operation of the wind turbine and feedback information. For example, the model can be retrained or fine-tuned using new data; or methods such as transfer learning or incremental learning can be used to enable the model to adapt to new data distribution or changes. These methods enable the model to stay up-to-date and optimal, and improve the accuracy and robustness of predictions.

Through the above methods, the life of the wind turbine can be predicted and optimized, and the performance degradation and life extension of the wind turbine can be dynamically adjusted, thereby improving the operating efficiency and economy of the wind farm.

It is obvious to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as illustrative and non-restrictive from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is therefore intended that all claims falling within the claims All changes within the meaning and scope of equivalent elements are included in the present invention. Any reference signs in the claims shall not be construed as limiting the claim in question.

In addition, it should be understood that although this specification is described in terms of implementations, not each implementation only contains an independent technical solution. This description of the specification is only for the sake of clarity, and those skilled in the art should take the specification as a whole. , the technical solutions in each embodiment can also be appropriately combined to form other implementations that can be understood by those skilled in the art.

Claims (10)

1. The multi-model fusion method for predicting the service life of the wind turbine generator is characterized by comprising the following steps of:

s101, data collection and preprocessing

Collecting data from various data sources of the running state, the environmental parameters and the fault records of the wind turbine generator, and performing pretreatment operations of data cleaning, data normalization and feature extraction to obtain a data set suitable for model input;

s102, model construction and model training

Selecting a proper machine learning model, namely model selection, according to different data characteristics and prediction targets, and respectively carrying out model training on a data set to obtain a plurality of sub-models;

s103, model fusion and model evaluation

Adopting one or more model fusion methods to synthesize the outputs of a plurality of sub-models to obtain a final life prediction result, and evaluating the performance and accuracy of the models by the methods of cross verification and error analysis;

s104, model application and model update

And deploying the model into a monitoring system of the wind farm, predicting the service life of the wind turbine at regular intervals, and adjusting and updating the model according to actual running conditions and feedback information.

2. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the data collection step, various data are acquired from the wind turbine generator in real time or periodically by using a mode of a sensor, a monitoring system and a log file, and are stored in a database or a cloud platform.

3. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the data cleaning step, the quality inspection and correction are carried out on the data by using an anomaly detection method, a missing value filling method and a noise filtering method, and incomplete, inaccurate or irrelevant data are removed or corrected.

4. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the data normalization step, the data are subjected to scale adjustment and distribution transformation by using maximum and minimum value normalization, standardization and logarithmic transformation methods, and the dimensional and deviation effects of the data are eliminated.

5. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the feature extraction step, the data is subjected to dimension reduction and transformation by using principal component analysis, factor analysis and wavelet transformation methods, and main features and information of the data are extracted.

6. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the model selection step, an appropriate machine learning model is selected based on the type and distribution of data, and predicted tasks and metrics.

7. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the model training step, a training set, a verification set, a corresponding loss function and an optimization algorithm are used for carrying out parameter initialization and iterative updating on the model, and a regularization and early stop method is used for preventing the model from being over-fitted or under-fitted.

8. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the model fusion step, a proper model fusion method is selected according to the type and output of the submodels, and the complementarity and diversity among the submodels are utilized to improve the stability and reliability of life prediction.

9. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the model evaluation step, the prediction result of the model is evaluated using the test set or new data and the corresponding evaluation index and method, and the performance and accuracy of the model are checked.

10. The multi-model fusion method for predicting service life of wind turbine generator set according to claim 1, wherein the method comprises the following steps: in the model updating step, the model is adjusted and updated according to the actual running condition and feedback information of the wind turbine generator.