Wind driven generator blade icing fault prediction method based on deep learning

Abstract

The invention relates to a wind driven generator blade icing fault prediction method based on deep learning, and the method comprises the following steps: constructing an integrated learning classifier XGBoost which is a gradient lifting integrated algorithm based on a decision tree, and gives consideration to a linear model solver and a decision tree learning algorithm; deep feature learning: through the ability of deep feature learning, mining the internal relation between the operation parameters and the blade health state, establishing a multi-layer deep network structure, deeply learning the nonlinear high-order correlation between the operation parameters of the wind driven generator and an output target, and completing the complex mapping from each working condition parameter to hidden layer feature output; and carrying out classification mode identification of an integrated classifier, taking hidden layer output information of deep network learning under different dimensions as input of an integrated learning classifier XGBoost, obtaining evaluation indexes under different feature dimensions through weighted classification accuracy of a plurality of base classifiers, and carrying out prediction on the icing fault of the wind driven generator blade.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine fault prediction, and in particular to a wind turbine blade icing fault prediction method based on deep learning. Background Art

[0002] With the increasing consumption of non-renewable resources such as oil, natural gas, and coal, environmental pollution is becoming more and more serious, and the energy crisis is becoming more and more prominent; countries around the world pay more attention to research and investment in energy, especially renewable resources; wind energy resources are a renewable and clean energy with large reserves and convenient utilization. Its development and utilization has become an important area that countries around the world are competing to develop; the main way to utilize wind energy resources is wind power generation, which inevitably requires the use of wind turbines.

[0003] The surface of wind turbine blades placed in extremely cold and low-temperature areas will inevitably face the problem of icing, which poses a great threat to the power generation performance and safe operation of wind turbines. With the continuous increase in the design power of wind turbines, the height of existing wind turbine towers continues to grow. Even in the northern coastal and mountainous areas, a large number of wind turbines will touch the lower clouds in winter, and they are very easy to freeze in low temperature and humid environments. Although many new wind turbines are designed with automatic de-icing systems, the challenge faced in practical applications is that it is difficult to accurately predict the early process of icing so that the de-icing system can be activated as early as possible. When the alarm is triggered, the blades are often already frozen over a large area. Operating under such conditions will increase the risk of blade breakage and damage.

[0004] In recent years, some research units have been committed to developing blade icing prediction models suitable for a certain region, mainly tending to study four aspects: icing meteorological conditions, ice accumulation model simulation, prediction of the impact of icing on power and icing power; meteorological icing is mainly related to liquid water content, air temperature and wind speed; by analyzing ground meteorological observation data and numerical weather forecast data and using statistical methods, the icing climate conditions are inferred, but this detection method has regional limitations; ice accumulation model simulation is not only related to meteorology, but also to the shape and properties of the blades. Through wind tunnel experimental research on blade ice accumulation simulation, the icing event is realized. Fault detection, but this ice accumulation model detection method does not take into account meteorological conditions or mechanical factors; icing events will affect the power generation of wind turbines. By coupling the icing model and the numerical weather model, the icing condition of the blades is monitored with reference to the wind turbine power curve, but the technology for predicting the power generation of wind farms in cold climates is not currently operational; the difference in power under icing conditions and non-icing conditions is compared to predict the icing condition of the blades; however, this prediction method has great uncertainty because of the lack of verified ice load data and analysis of the freezing development process, and it is impossible to completely determine that the reduction in power generation is caused by icing.

[0005] Through analysis and comparison, the above methods all have certain limitations. The SCADA system (Supervisory Control and Data Acquisition) is an important system for wind farm equipment management, monitoring and control, which stores real-time detection data of wind turbine operation; these real-time data fully cover the external environment information and internal working status of wind turbines. If the deep learning model can be used to mine the blade icing status of fault information from the SCADA data, and complete the blade icing detection and prediction work, it will provide technical support for the early activation of the de-icing system of the blade.

[0006] During operation, the normal operation time of the blades is often longer than the ice-covered time, and the generated data often has the problem of class imbalance; undersampling of normal samples and oversampling of faulty samples are the two main methods to solve this problem; however, these two methods often have the following problems: the new samples generated by oversampling are prone to noise and even lead to overfitting, and undersampling has the problem of losing important information; in addition, the ratio of undersampling and oversampling is often determined by trial and error, which increases the complexity of the test process. Summary of the invention

[0007] The purpose of the present invention is to provide a wind turbine blade icing fault prediction method based on deep learning, which uses the excellent adaptive feature extraction capability of the deep network to perform high-dimensional feature representation on the original unbalanced data, and uses the strong classification capability of the integrated classifier to identify the extracted features, thereby realizing wind turbine blade icing fault prediction.

[0008] In order to achieve the above objectives, the technical solution adopted by the present invention is: a method for predicting icing faults of wind turbine blades based on deep learning, characterized in that it includes the following steps:

[0009] Step 1: Build an ensemble learning classifier XGBoost. XGBoost is a gradient boosting ensemble algorithm based on decision trees, which combines linear model solvers and decision tree learning algorithms. It combines many decision tree models to form a model with high accuracy.

[0010] Step 2: Deep feature learning. The operating parameters monitored by SCADA are selected as the features for predicting blade icing faults. Through the ability of deep feature learning, the intrinsic relationship between the operating parameters and the health status of the blades is explored, and a multi-layer deep network structure is established to deeply learn the nonlinear high-order correlation between the operating parameters of the wind turbine and the output target, completing the complex mapping of various operating parameters to the hidden layer feature output;

[0011] Step 3: Classification pattern recognition by integrated classifier: the hidden layer output information under different dimensions of deep network learning is used as the input of the integrated learning classifier XGBoost. The evaluation index under different feature dimensions is obtained through the weighted classification accuracy of multiple base classifiers to predict the icing failure of wind turbine blades.

[0012] Preferably, the specific operation of XGBoost is as follows: assuming that a sample set with n samples and m features is given, that is, D = {xi, yi}, where x and y represent the feature vector and the corresponding dependent variable respectively; during the training process, an objective function is set:

[0013] obj(w)=l(w)+Ω(w)

[0014] Among them, w is the parameter to be estimated in the model, l(w) is the loss function of the model, and Ω(w) is the regularization term, which is used to control the complexity of the model;

[0015] The loss function can be expressed as:

[0016]

[0017]

[0018] in is the model prediction value in round t, fk(xi) represents the regression decision tree;

[0019] The objective function can be reformulated as:

[0020]

[0021] Take the form of Taylor quadratic expansion

[0022]

[0023] And define

[0024]

[0025] To approximate the objective function, it is convenient for further calculation; objective function:

[0026]

[0027] For the convenience of research, the constant term is removed, and the objective function becomes

[0028]

[0029] Split the decision tree into the tree structure part q and the weight (score) part ω of the leaf node, then each tree can be expressed as

[0030] f t (x) = ω q(x)

[0031] Redefine the complexity of the tree:

[0032]

[0033] Under these new definitions, the new form of the objective function is

[0034]

[0035] This objective consists of T mutually independent quadratic functions of a single variable, which can be defined as:

[0036]

[0037] After the objective function is introduced, its relationship with the variable ω j Taking the derivative, we get ω j The optimal value of:

[0038]

[0039] Substituting it into the objective function, we get:

[0040]

[0041] This will give you the best decision tree

[0042]

[0043] XGBoost uses a greedy search algorithm to find the best tree structure; the algorithm is the so-called "exact greedy split algorithm"; XGBoost starts from the root node and iteratively splits the features, assuming that IL and IR represent the sample sets of the left and right branches respectively; for each node on the tree, XGBoost will try to add a split; as the objective function changes, the information gain obtained after adding the split point is described as follows:

[0044]

[0045] The above formula contains four scores, namely the score of the left subtree, the score of the right subtree, the score when the node is not split, and γ represents the complexity added when adding a new child node; if Gain<0, XGBoost stops splitting, which means that the parameter γ handles the complexity here; the split point with the maximum value of Gain is considered to be the optimal split on the node;

[0046] XGBoost uses learning rate, boosting number, maximum tree depth and subsampling to avoid overfitting of the model. In addition, there are three other hyperparameters in XGBoost: γ∈(0,+∞) specifies the minimum loss function drop value required for node splitting; when a node is split, the node will be split only if the loss function value after the split decreases; ω mc ∈(0,+∞) is the minimum sample weight among the child nodes. If the sum of the sample weights of a leaf node is less than ω mc , the splitting process ends; in addition, XGBoost also uses column (feature) sampling (Column Subsampling) technology, which prevents overfitting better than traditional row sampling and is conducive to parallel processing of data; hyperparameters enable XGBoost to randomly collect some features in the growing tree, thereby accelerating the learning speed and avoiding overfitting.

[0047] The technical effect of the present invention is as follows: in order to solve these problems, firstly, the original unbalanced data is represented by high-dimensional features using the excellent adaptive feature extraction capability of the deep network, and then the extracted features are used for fault identification using the strong classification capability of the integrated classifier; the model is innovatively applied to blade icing fault identification to achieve wind turbine blade icing fault prediction. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below. The description in this section is only exemplary and explanatory and should not have any limiting effect on the scope of protection of the present invention.

[0049] This application discloses a method for predicting wind turbine blade icing failure based on deep learning. The specific method is as follows:

[0050] 1. Extreme Gradient Boosting Decision Tree

[0051] Extreme Gradient Boosting (XGBoost) is a gradient boosting ensemble algorithm based on decision trees, combining linear model solvers and decision tree learning algorithms. Its basic idea is to combine many decision tree models to form a highly accurate model. Its fast, efficient, and scalable system has achieved promising results on many standard classification benchmarks.

[0052] Assume that there are n sample sets and m features, that is, D = {x i ,y i}, where x and y represent the feature vector and the corresponding dependent variable respectively; during the training process, an objective function is set:

[0053] obj(w)=l(w)+Ω(w),(3.1)

[0054] Among them, w is the parameter to be estimated in the model, l(w) is the loss function of the model, and Ω(w) is the regularization term, which is used to control the complexity of the model;

[0055] The loss function can be expressed as:

[0056]

[0057] in is the model prediction value in round t, f k (x i ) represents a regression decision tree.

[0058] The objective function can be reformulated as:

[0059]

[0060] Take the form of Taylor quadratic expansion

[0061]

[0062] And define

[0063]

[0064] To approximate the objective function, it is convenient for further calculation. Objective function:

[0065]

[0066] For the convenience of research, the constant term is removed, and the objective function becomes

[0067]

[0068] Split the decision tree into the tree structure part q and the weight (score) part ω of the leaf node, then each tree can be expressed as

[0069] f t (x) = ω q(x) ,(3.8)

[0070] Redefine the complexity of the tree:

[0071]

[0072] Under these new definitions, the new form of the objective function is

[0073]

[0074] This objective consists of T independent quadratic functions of a single variable; we can define:

[0075]

[0076] After the objective function is introduced, its relationship with the variable ω j Taking the derivative, we get ω j The optimal value of:

[0077]

[0078] Substituting it into the objective function, we get:

[0079]

[0080] This will give you the best decision tree

[0081]

[0082] XGBoost uses a greedy search algorithm to find the best tree structure. The algorithm is the so-called "exact greedy split algorithm"; XGBoost starts from the root node and iteratively splits the features. Assume that IL and IR represent the sample sets of the left and right branches respectively. For each node on the tree, XGBoost will try to add a split; as the objective function changes, the information gain obtained after adding the split point is described as follows:

[0083]

[0084] Formula (3.15) contains four scores, namely the score of the left subtree, the score of the right subtree, the score when the node is not split, and γ represents the complexity added when adding a new child node. If Gain<0, XGBoost stops splitting, which means that the parameter γ handles complexity here; the split point with the maximum value of Gain is considered to be the optimal split on the node.

[0085] XGBoost uses learning rate, boosting number, maximum tree depth and subsampling to avoid overfitting of the model. In addition, there are three other hyperparameters in XGBoost: γ∈(0,+∞) specifies the minimum loss function drop required for node splitting; when a node is split, the node will be split only if the loss function value after the split decreases; ω mc∈(0,+∞) is the minimum sample weight in the child node. If the sum of the sample weights of a leaf node is less than ω mc , the splitting process ends; in addition, XGBoost also uses column (feature) sampling (Column Subsampling) technology, which prevents overfitting better than traditional row sampling and is conducive to parallel processing of data; hyperparameters enable XGBoost to randomly collect some features in the growing tree, thereby accelerating the learning speed and avoiding overfitting.

[0086] 2. Establishment of fault prediction model

[0087] When wind turbine blades are frozen, their operating monitoring parameters often change significantly; for example, during the icing period, ice covers the rotor blades, reducing aerodynamic efficiency and torque, resulting in increased power loss, etc. Therefore, the operating parameters monitored by SCADA can be used as features for blade icing fault prediction (specifically wind speed, generator speed, grid-side active power (kw), wind angle (°), 25-second average wind direction angle, yaw position, yaw speed, blade angle, blade speed, pitch motor temperature, xy acceleration, ambient temperature, cabin temperature, charger DC current, etc.), and through the ability of deep feature learning, the intrinsic relationship between operating parameters and blade health status can be explored, and the classification ability of the classifier can be used to accurately locate the icing status. The detailed steps are as follows:

[0088] Step 1: Deep feature learning. Through the established multi-layer deep network structure, the nonlinear high-order correlation between the operating parameters of the wind turbine and the output target is deeply learned, and the complex mapping of various operating parameters to the hidden layer feature output is completed.

[0089] Step 2: Classification pattern recognition by ensemble classifier: the hidden layer output information under different dimensions of deep network learning is used as the input of the ensemble learning classifier XGBoost. The evaluation index under different feature dimensions is obtained through the weighted classification accuracy of multiple base classifiers.

[0090] 3. Evaluation indicators

[0091] Given that the influence of large classes on the accuracy standard is greater than that of rare classes, making it difficult to improve the recognition rate of rare classes, the new classifier evaluation index focuses more on the influence of rare classes on performance indicators. At this time, the accuracy cannot fully explain the classification ability of the model, so the area under the receiver operating characteristic (ROC) curve (Area Under the ROC Curve, AUC) is used as the classifier evaluation index for class imbalanced samples; AUC refers to the area under the ROC curve, that is, the integral of the ROC curve. The larger the AUC area value, the stronger the classification ability of the classifier.

[0092] From the binary confusion matrix shown in Table 3.1, we can calculate precision = TP / TP+FP, recall = TP / TP+FN; accuracy = (TP+TN) / (TP+TN+FP+FN); the X-axis of the ROC curve represents the true positive rate (True Positive Rate, TPR), and the Y-axis represents the false positive rate (False Positive Rate, FPR). The set of points on it is obtained by adjusting the decision threshold of the classifier.

[0093] Table 3.1 Confusion matrix for binary classification

[0094] Model prediction value: 0 Model prediction value: 1 True value: 0 TruePositive(TP) FalseNegative(FN) True value: 1 FalsePositive(FP) TrueNegative(TN)

[0095]

[0096] In practical applications, the data monitored by the SCADA system can be sampled at intervals as features for predicting blade icing faults (specifically wind speed, generator speed, grid-side active power (kW), wind angle (°), 25-second average wind direction angle, yaw position, yaw speed, blade angle, blade speed, pitch motor temperature, xy direction acceleration, ambient temperature, cabin temperature, charger DC current, etc.), and then the sampled data can be applied to the icing fault prediction model of the present application for calculation, so as to realize the prediction of icing faults of wind turbine blades.

[0097] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0098] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. The above is only a preferred implementation of the present invention. It should be pointed out that due to the limitations of textual expression and the objective existence of infinite specific structures, ordinary technicians in this technical field can make several improvements, modifications or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the protection scope of the present invention.

Claims

1. A wind turbine blade icing fault prediction method based on deep learning, characterized in that: The steps include: Step 1: Build an ensemble learning classifier XGBoost. XGBoost is a gradient boosting ensemble algorithm based on decision trees, which combines linear model solvers and decision tree learning algorithms. It combines many decision tree models to form a model with high accuracy. Step 2: Deep feature learning. The operating parameters monitored by SCADA are selected as the features for predicting blade icing faults. Through the ability of deep feature learning, the intrinsic relationship between the operating parameters and the health status of the blades is explored, and a multi-layer deep network structure is established to deeply learn the nonlinear high-order correlation between the operating parameters of the wind turbine and the output target, completing the complex mapping of various operating parameters to the hidden layer feature output; Step 3: Classification pattern recognition by integrated classifier: the hidden layer output information under different dimensions of deep network learning is used as the input of the integrated learning classifier XGBoost. The evaluation index under different feature dimensions is obtained through the weighted classification accuracy of multiple base classifiers to predict the icing failure of wind turbine blades.

2. The method for predicting wind turbine blade icing faults based on deep learning according to claim 1, characterized in that: The specific operation of XGBoost is as follows: Assume that a sample set with n samples and m features is given, that is, D = {xi, yi}, where x and y represent the feature vector and the corresponding dependent variable respectively; during the training process, set an objective function: obj(w)=l(w)+Ω(w) Among them, w is the parameter to be estimated in the model, l(w) is the loss function of the model, and Ω(w) is the regularization term, which is used to control the complexity of the model; The loss function can be expressed as: in is the model prediction value in round t, fk(xi) represents the regression decision tree; The objective function can be reformulated as: Take the form of Taylor quadratic expansion And define To approximate the objective function, it is convenient for further calculation; objective function: For the convenience of research, the constant term is removed, and the objective function becomes Split the decision tree into the tree structure part q and the weight (score) part ω of the leaf node, then each tree can be expressed as f t (x)=ω q(x) Redefine the complexity of the tree: Under these new definitions, the new form of the objective function is This objective consists of T mutually independent quadratic functions of a single variable, which can be defined as: After the objective function is introduced, its relationship with the variable ω j Taking the derivative, we get ω j The optimal value of: Substituting it into the objective function, we get: This will give you the best decision tree XGBoost uses a greedy search algorithm to find the best tree structure; the algorithm is the so-called "exact greedy split algorithm"; XGBoost starts from the root node and iteratively splits the features, assuming that IL and IR represent the sample sets of the left and right branches respectively; for each node on the tree, XGBoost will try to add a split; as the objective function changes, the information gain obtained after adding the split point is described as follows: The above formula contains four scores, namely the score of the left subtree, the score of the right subtree, the score when the node is not split, and γ represents the complexity added when adding a new child node; if Gain<0, XGBoost stops splitting, which means that the parameter γ handles the complexity here; the split point with the maximum value of Gain is considered to be the optimal split on the node; XGBoost uses learning rate, boosting number, maximum tree depth and subsampling to avoid overfitting of the model. In addition, there are three other hyperparameters in XGBoost: γ∈(0,+∞) specifies the minimum loss function drop required for node splitting; when a node is split, the node will be split only if the loss function value after the split decreases; ω mc ∈(0,+∞) is the minimum sample weight in the child node. If the sum of the sample weights of a leaf node is less than ω mc , the splitting process ends; in addition, XGBoost also uses column (feature) sampling (Column Subsampling) technology, which prevents overfitting better than traditional row sampling and is conducive to parallel processing of data; hyperparameters enable XGBoost to randomly collect some features in the growing tree, thereby accelerating the learning speed and avoiding overfitting.

Images