A comprehensive evaluation method for wind turbine performance under normal and extreme conditions
Through one-dimensional residual neural network and XGBoost algorithm, combined with the whole machine simulation model and data preprocessing, key features are selected for wind turbine performance evaluation, which solves the problem that the existing technology is difficult to evaluate the overall performance of the wind turbine, and achieves a higher level of assessment accuracy and efficiency.
Patent Information
- Application Number
- CN202311060758.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-22
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2043-08-22
AI Technical Summary
The existing technology is difficult to achieve comprehensive performance evaluation of the wind turbine unit as a whole, and cannot meet the performance evaluation needs of wind turbine units under extreme operating conditions such as complex aerodynamics, hydropower and mooring anchor chain loads, resulting in high maintenance costs and low power generation benefits.
Using one-dimensional residual neural network combined with XGBoost algorithm, through the whole machine simulation model and data preprocessing, key features are selected for comprehensive performance analysis, and the network is deepened by one-dimensional convolutional layer and residual blocks, improving the evaluation accuracy and generalization of the model ability to transform.
It improves the accuracy and efficiency of the overall performance evaluation of wind turbine units, reduces the evaluation cost, enhances the generalization ability of the model, solves the shortcomings of the single component evaluation method, and is suitable for wind turbine performance evaluation under normal and extreme operating conditions.
Smart Images

Figure CN117333029B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wind power generation, and in particular to a method for comprehensively evaluating the overall performance of a wind turbine generator set under normal and extreme working conditions. Background Art
[0002] In order to achieve the "dual carbon" goal and sustainable social development, my country has been actively developing environmentally friendly, clean and pollution-free renewable energy in recent years. Wind energy has become one of the most promising green energy sources due to its advantages of being clean and pollution-free, widely distributed, with huge reserves and unlimited renewable energy. As the main equipment for wind power generation, wind turbines convert wind energy into high-speed, low-torque mechanical energy to drive generators to generate electricity. Wind turbines have complex structures and are subjected to disordered and time-varying alternating loads for a long time. In addition, the installation location is relatively remote and the maintenance cost is high. Therefore, in order to ensure the normal operation of wind turbines, reduce maintenance costs and increase power generation revenue, it is very important to find a comprehensive evaluation method for the performance of wind turbines under normal and extreme conditions.
[0003] Previous evaluation methods often focus on the performance evaluation of single components such as blades, gearboxes, towers, etc. of wind turbines, and cannot achieve the comprehensive performance evaluation of the entire wind turbine. As wind turbines move from land to sea, and from offshore fixed wind turbines to offshore wind turbines, they are subject to greater and more complex aerodynamic forces, hydrodynamic forces, mooring chain loads, etc., and the comprehensive performance requirements of the entire machine are getting higher and higher. The previous single component evaluation method can no longer meet the requirements. This paper adopts the one-dimensional residual neural network method, considering the five characteristics of wind turbine reliability, stability, response characteristics, power characteristics, and power generation characteristics, to comprehensively evaluate the comprehensive performance of the entire wind turbine. Summary of the invention
[0004] The present invention provides a comprehensive evaluation method for the overall performance of a wind turbine under normal and extreme working conditions. By comprehensively analyzing the overall performance of the wind turbine and visualizing the data, rich features are established. Feature selection is performed based on the XGBoost algorithm to mine feature data that better characterizes the overall performance of the whole machine, thereby reducing the shortcomings of manually selected features and enhancing the evaluation performance and generalization ability of the model. The method is simple to implement and has low cost.
[0005] In order to achieve the above technical features, the purpose of the present invention is achieved as follows: a method for comprehensive evaluation of wind turbine performance under normal and extreme conditions, comprising the following steps:
[0006] S1, determine the whole machine parameters according to the machine model selection, installation location, design requirements, installation and maintenance cost requirements;
[0007] S2, based on the measured data of the wind field at the installation location, calculate and select the distribution function and model of the average wind speed and the pulsating wind speed under normal and extreme working conditions;
[0008] S3, build a wind turbine simulation model based on FAST, and solve the model to obtain the required data;
[0009] S4, performing a series of preprocessing operations such as data labeling, data resampling and data normalization on the acquired data;
[0010] S5, performs comprehensive performance analysis and XGBoost feature selection based on the whole machine on the preprocessed data;
[0011] S6, selecting a corresponding reference value, comparing the data obtained by solving the model with the corresponding reference value, and judging the degree of excellence of the overall performance of the whole machine based on the data;
[0012] S7, assign weights to each feature according to design requirements and engineering experience, sort the features by importance, and use the method of increasing feature dimensions to compare the accuracy of the one-dimensional residual neural network model;
[0013] S8, using the test set to test and evaluate the comprehensive performance of the wind turbine model.
[0014] The overall parameters of a wind turbine include: rated power, rotor layout, rotor configuration, control method, total gearbox ratio, cut-in, rated and cut-out wind speeds.
[0015] The comprehensive performance characteristics of the whole machine include: reliability, stability, response characteristics, power characteristics and power generation characteristics.
[0016] The modules required to construct the whole machine model based on FAST include: pneumatic, hydrodynamic, structural, servo and mooring modules.
[0017] Preprocess the data obtained from S3. The specific steps are as follows:
[0018] S3.1, mark the acquired data and remove invalid data;
[0019] S3.2, re-dividing the data by resampling, resampling the original data set every ten minutes, and taking the average value of each group of sample values as the new sample feature;
[0020] S3.3, through normalization processing to eliminate the range differences between the values of different features, improve the learning efficiency of the model.
[0021] In S5, relevant features are supplemented based on the overall performance analysis of the whole machine, and finally feature selection is performed through the XGBoost algorithm.
[0022] In summary, the present invention performs a coarse filtering of the original features by analyzing the features selected and generated by the whole machine data, and then assigns weights to the feature indicators according to the design requirements and engineering experience, and selects the comprehensive performance features of the whole machine by sorting the feature importance through XGBoost, thereby enhancing the accuracy and versatility of the model. A one-dimensional residual neural network is used to adaptively extract effective features and make predictions, and the effectiveness and accuracy of the proposed method are verified using the whole machine data from software simulation. The one-dimensional convolutional layer adaptively extracts features and makes a comprehensive performance evaluation of the whole machine. The addition of residual blocks deepens the network while solving the problems of gradient vanishing and gradient explosion in the process of network optimization, and has high comprehensive performance and generalization ability. This method can utilize the powerful learning ability of deep learning to obtain the best classification effect, has high prediction accuracy and generalization ability, and can increase the efficiency of comprehensive performance evaluation to the maximum value.
[0023] The comprehensive evaluation method of wind turbine performance under normal and extreme working conditions of the present invention has the following beneficial effects:
[0024] 1. By analyzing the comprehensive performance of the wind turbine set and selecting features based on the XGBoost algorithm as the input features of the training model, compared with the original features, 1DRes-CNN has better accuracy in comprehensive performance evaluation of the whole machine based on the features selected by XGBoost.
[0025] 2. Apply the one-dimensional convolutional residual neural network to the comprehensive performance evaluation of the wind turbine set. Use the one-dimensional convolutional layer to adaptively extract features and make a comprehensive performance evaluation of the whole machine. Add a one-dimensional residual block to deepen the network while solving the problems of gradient vanishing and gradient explosion in the network optimization process.
[0026] 3. Reduce the cost of comprehensive performance evaluation of wind turbines and the complexity of implementation methods.
[0027] 4. Reduce the shortcomings of artificial feature selection in feature selection, while enhancing the evaluation performance and generalization ability of the model.
[0028] 5. Reduce the problem of DNN training difficulties and performance degradation caused by the increase of network depth during network optimization. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0030] Figure 1 The present invention is a flow chart of a method for comprehensively evaluating the overall performance of a wind turbine generator set under normal and extreme working conditions.
[0031] Figure 2 It is the one-dimensional residual neural network of the present invention.
[0032] Figure 3 It is a flow chart of the overall idea of the present invention. DETAILED DESCRIPTION
[0033] The embodiments of the present invention are further described below in conjunction with the accompanying drawings.
[0034] S1, determination of whole machine parameters. The whole machine parameters are determined according to the machine type selection, installation location, whole machine power and economy, installation and maintenance costs, etc. The whole machine parameters of the wind turbine set include: rated power, wind rotor layout, wind rotor configuration, control method, total gearbox transmission ratio, cut-in, rated, cut-out wind speed, etc.
[0035] S2, selection of average wind speed. Wind speed frequency is an important parameter that reflects the statistical characteristics of wind. Usually, the percentage of the duration of the same wind speed in a certain period of time to the total measurement duration is called wind speed frequency. The models usually used to fit the wind speed frequency distribution include Rayleigh distribution, logarithmic distribution, Г distribution, two-parameter Weibull distribution, three-parameter Weibull distribution and Pearson distribution. Among them, Rayleigh distribution and two-parameter Weibull distribution are simple probability models that can better fit the actual wind speed frequency distribution.
[0036] Weibull distribution probability density function:
[0037]
[0038] Distribution function:
[0039]
[0040] Where v represents the 10-minute average wind speed; k and c are the shape parameter and size parameter of the Weibull distribution, respectively. When k = 2, it becomes the Rayleigh distribution.
[0041] In IEC 61400-1, the 10-minute average wind speed at hub height is assumed to conform to the following Rayleigh distribution:
[0042]
[0043] In the formula, v hub Indicates the 10-minute average wind speed at the hub height, v ave Represents the annual average wind speed.
[0044] In IEC 61400-1, the average wind speed variation with height should be given by the exponential:
[0045]
[0046] In the formula, z (hub) It represents the height from the ground at the wheel hub, z represents the desired height from the ground, and α represents the wind shear coefficient, which is an empirical value.
[0047] S3, selection of fluctuating wind speed. The contribution of kinetic energy of eddies of different sizes to the kinetic energy of fluctuating wind can be described by turbulence power spectrum density. Commonly used longitudinal turbulence power spectra include Davenport spectrum, Harris spectrum, Kaimal spectrum, Karman spectrum and Simiu spectrum.
[0048] In IEC 61400-1, two turbulence models are given for generating fluctuating wind: the Mann turbulence model and the Kaimal turbulence model. For the Kaimal turbulence model:
[0049]
[0050] Where k is the velocity component subscript, indicating the direction (u is longitudinal, v is transverse, w is vertical); S k represents the power spectrum of the k velocity component; σ k represents the standard deviation of the k velocity component; L k represents the integral scale parameter of the k velocity component; v hub It represents the 10-minute average wind speed at the hub height; f represents the frequency.
[0051] The Kaimal power spectrum can be used with the following exponential coherence function to calculate the longitudinal velocity component at each point in space (the Kaimal turbulence model does not define transverse and vertical coherence functions):
[0052]
[0053] In the formula, r represents the distance between any two points in space; f represents the frequency; v hub Indicates the wind speed at the hub; L u Represents the integral scale parameter in the u direction (vertical direction).
[0054] S4, wind turbine complete machine simulation model construction and solution. Use FAST to build the wind turbine complete machine model, and solve it to obtain the required data. The modules required to build the complete machine model include: pneumatic, hydrodynamic, structural, servo, and mooring modules.
[0055] S5, preprocess the acquired data, including: data labeling, resampling, and data normalization. Divide the data set into a training set and a validation set. First, label the wind turbine data and remove invalid data; then resample the data and re-divide the data. Resample the data every ten minutes and take the average value of each group of sample values as the new sample feature; normalize the data to eliminate the range differences between the values of different features and improve the learning efficiency of the model. The normalization formula is:
[0056]
[0057] In the formula, X represents the original data in the dataset; X (norm) Indicates the normalized data; X max Indicates the largest data in the data set; X min Represents the smallest data in the data set.
[0058] S6, after data preprocessing, conducts feature engineering for comprehensive performance analysis of the wind turbine set and XGBoost feature importance calculation: first, analyze the features related to the comprehensive performance of the wind turbine set, and finally perform feature selection through the XGBoost algorithm.
[0059] S6.1, Wind turbines are now developing in the direction of complexity and large-scale. In a variety of complex and harsh operating environments such as complex and changeable wind conditions, high temperature, high humidity, etc., large wind turbines face severe reliability and stability issues. Wind turbines have high investment risks and high operation and maintenance costs, which put forward higher requirements on the reliability and stability of wind turbines. Improving reliability and stability is the cornerstone of promoting the development of wind turbines. Therefore, reliability and stability are important indicators for evaluating the comprehensive performance of the whole machine. The new characteristic value is recorded as: v (r) 、v (s) .
[0060] S6.2, under various working conditions, the accurate calculation of the dynamic response of the whole unit is the fundamental guarantee for the safe operation of the unit, and it can also effectively reduce the value of the safety factor in the design to make the unit have better economic benefits. Therefore, the response characteristic is an important indicator for evaluating the comprehensive performance of the whole unit. The new characteristic value is recorded as: v (re) .
[0061] S6.3, power characteristics are an important indicator for the overall performance evaluation of wind turbines. Power characteristics are the wind speed v (i) As the horizontal axis, the active power p (i) A series of normalized data pairs (v (i) , p (i) ) is a characteristic curve drawn by . It can be seen that the power characteristic is a description of the power generation capacity of the wind turbine group. The quality of the power characteristic will directly affect the power generation of the wind turbine group. Therefore, the power characteristic is an important indicator for evaluating the comprehensive performance of the whole machine. The new characteristic value is recorded as: v (p) .
[0062] S6.4, wind turbines use wind energy to generate electricity, and wind turbines are devices that convert wind energy into electrical energy. Wind energy is converted into mechanical energy through blades, and then mechanical energy is converted into electrical energy through generators. The power generation performance of wind turbines directly determines their economic benefits. If the benefits are too low, it will directly affect the development of the wind turbine industry. Therefore, the power generation characteristics are an important indicator for evaluating the overall performance of the whole machine. The new characteristic value is recorded as: v (G) .
[0063] S7, feature selection through XGBoost. XGBoost can be used for feature importance selection. When the algorithm constructs the boosted tree design, the importance of each feature to the training model is indicated by obtaining a feature score. The more a feature is used to enhance the key decision of the tree, the higher its score. The features generated by the performance analysis of the wind turbine and the original features are used as the input of the XGBoost algorithm to obtain the feature importance of the data. The importance is the number of times the feature partitions the tree nodes. Each time the partition is performed, the feature with the largest information gain is selected as the split point by the greedy method. The information gain is calculated as follows:
[0064]
[0065] In the formula, g l represents the first-order derivative of the loss function; h l represents the second-order derivative of the loss function; g r represents the first-order derivative of the information gain rate; h r represents the second-order derivative of the information gain rate; λ represents L 2 The coefficient of the norm; γ represents the shrinkage coefficient.
[0066] S8, selection of relative reference values. When evaluating the performance of the whole machine, the corresponding reference values of the required characteristics come from: if the performance evaluation of the whole machine is applicable to the whole machine design process, then the whole machine data of typical wind turbines of the same or similar type are selected from the publicly available wind turbine data as the reference value; if there is no similar wind turbine data, then the data of the given target wind turbine is used as the reference value, which depends on the standards, codes and experience of designing wind turbines.
[0067] S9, after obtaining the relative reference values of various characteristics of the wind turbine generator set, compare various characteristic data of the wind turbine generator set to be evaluated with the corresponding reference values.
[0068] S10, the performance evaluation of wind turbines is divided into six levels: "excellent", "good", "qualified", "defective", "poor", and "useless". If 5 parameters are within the allowable range, it is judged as "excellent"; 4 parameters are within the allowable range, it is judged as "good"; 3 parameters are within the allowable range, it is judged as "qualified"; 2 parameters are within the allowable range, it is judged as "defective"; 1 parameter is within the allowable range, it is judged as "poor"; 0 parameters are within the allowable range, it is judged as "useless".
[0069] S11, construct a one-dimensional residual neural network (1DRes-CNN), add residual blocks to its neural network, each residual block consists of two convolutional layers, deepen the number of network layers to improve network efficiency while avoiding the loss of original features.
[0070] The structure of 1DRes-CNN is shown in the figure. The model input is the features selected by the XGBoost algorithm, and the output is 1, 0.8, 0.6, 0.4, 0.2, 0, where 1 means "excellent" overall performance; 0.8 means "good" overall performance; 0.6 means "qualified" overall performance; 0.4 means "defective" overall performance; 0.2 means "poor" overall performance; 0 means "useless" overall performance. In this model, the size of the convolution kernel is set to 3*1, and the number of convolution kernels is 16. The function of the Dropout layer is to randomly discard a part of neurons and connections according to a certain ratio, which is set to 0.25. The loss function selects the cross loss function.
[0071] S12, a one-dimensional convolutional layer is used to capture the local sequence features of the data information flow. The convolutional layer slides the convolution kernel on the input data to extract its local features. The neural network input data can be represented as x = [x1, x2..., xn], where n represents the length of the sequence. Therefore, the operation of the convolutional layer is:
[0072]
[0073] In the formula, represents the i-th feature of the l-th layer; f represents the activation function; represents the weight matrix; x l-1 Represents the output of the l-1th layer; Represents the bias term.
[0074] S13, residual network processing steps:
[0075] S13.1, input feature x 1 , after the first convolutional layer weighted and activation function operation in the residual module, the output F(x 1 ):
[0076] F(x 1)=f(ω 1 x 1 +b 1 )
[0077] Where f represents the ReLU activation function; ω 1 represents the weight of the first convolutional layer; b 1 Represents the bias of the first convolutional layer.
[0078] ReLU activation function f expression formula:
[0079] ψ(x)=max(0,x)
[0080] S13.2, F(x 1 ) Input the second convolutional layer to get the output formula:
[0081] F 1 (x 1 )=F(x 1 )ω 2 +b 2
[0082] In the formula, ω 2 represents the weight of the second convolutional layer; b 2 Represents the bias of the second convolutional layer.
[0083] S13.3, F 1 (x 1 ) and x 1 Input the ReLU activation function and get F 2 (x 1 ):
[0084] F 2 (x 1 )=f(F 1 (x 1 )+x 1 )
[0085] S14, network training uses the BP algorithm to reversely transfer errors, and then adjusts related parameters through the gradient descent method. The residual neural network will input x 1 Directly pass it to the output to fit the residual mapping function, and only learn the residual F between the input and output 1 (x 1 ), simplifying the learning objectives, so as to better exert the powerful learning ability of deep CNN.
[0086] S15, the comprehensive performance characteristics of the wind turbine include: reliability, stability, response characteristics, power characteristics, and power generation characteristics. Assign weights to the characteristic indicators according to design requirements and engineering experience. After sorting the feature importance, the accuracy of the 1DRes-CNN model diagnosis is compared in ascending feature dimensions. When the first 5 dimensional features of the features selected by XGBoost are selected, the 1DRes-CNN model has the highest classification accuracy. The 5 dimensional features are used as the input of the 1DRes-CNN model. The selected 5 dimensional features are shown in the following table:
[0087] Table 1. Characteristics of comprehensive evaluation of wind turbine performance
[0088] Serial number Parameter name Characteristic value name Weight 1 reliability <![CDATA[v (r) ]]> 20% 2 stability <![CDATA[v (s) ]]> 15% 3 Response characteristics <![CDATA[v (re) ]]> 25% 4 Power characteristics <![CDATA[v (p) ]]> 20% 5 Power generation characteristics <![CDATA[v (G) ]]> 20%
[0089] S16, using the test set to test and evaluate the overall performance of the wind turbine.
Claims
1. A comprehensive evaluation method for wind turbine performance under normal and extreme conditions, characterized in that: The following steps are involved: S1, determine the whole machine parameters according to the machine model selection, installation location, design requirements, installation and maintenance cost requirements; S2, based on the measured data of the wind field at the installation location, calculate and select the distribution function and model of the average wind speed and the fluctuating wind speed under normal and extreme working conditions; S3, build a wind turbine simulation model based on FAST, and solve the model to obtain the required data; S4, performing a series of preprocessing operations such as data labeling, data resampling and data normalization on the acquired data; S5, performs comprehensive performance analysis and XGBoost feature selection based on the whole machine on the preprocessed data; S6, selecting a corresponding reference value, comparing the data obtained by solving the model with the corresponding reference value, and judging the degree of excellence of the overall performance of the whole machine based on the data; S7, assign weights to each feature according to design requirements and engineering experience, sort the features by importance, and use the method of increasing feature dimensions to compare the accuracy of the one-dimensional residual neural network model; S8, using the test set to test and evaluate the comprehensive performance of the wind turbine model; Wind turbine parameters include: rated power, rotor layout, rotor configuration, control method, gearbox total transmission ratio, cut-in, rated and cut-out wind speeds; The comprehensive performance characteristics of the whole machine include: reliability, stability, response characteristics, power characteristics and power generation characteristics; The modules required to construct the whole machine model based on FAST include: pneumatic, hydrodynamic, structural, servo and mooring modules; Preprocess the data obtained from S3. The specific steps are as follows: S3.1, mark the acquired data and remove invalid data; S3.2, re-dividing the data by resampling, resampling the original data set every ten minutes, and taking the average value of each group of sample values as the new sample feature; S3.3, through normalization processing to eliminate the range differences between the values of different features, improve the learning efficiency of the model; In S5, relevant features are supplemented based on the overall performance analysis of the whole machine, and finally feature selection is performed through the XGBoost algorithm.
