Frequency variation based wind farm fault detection method

Abstract

The present application relates to a kind of wind farm fault detection method based on power system power characteristics.The method is by analyzing and determining the time series characteristics of power system frequency fluctuation under different wind turbine penetration, the variation law of analysis power spectrum characteristics with wind turbine penetration and the variation characteristics of frequency power spectrum density under different fault scenarios are obtained;Then the frequency of power system is monitored in real time, the time series of the frequency that changes is collected and its power spectrum is obtained;Finally, the time series and power spectrum obtained are compared by feature extraction and analysis, to determine which type of fault is more likely to occur in power system and make alarm.The method considers the different characteristics of frequency in time domain and frequency domain under different wind turbine penetration when different faults occur, has the advantages of high sensitivity, strong anti-interference ability and other methods, to a certain extent, eliminate the influence of noise and interference, improve the accuracy and reliability of fault detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system fault diagnosis, in particular to a wind farm fault detection method based on frequency change. BACKGROUND

[0002] With the increasing depletion of non-renewable resources such as coal and oil and the increasing demand for energy by the power system, the development of natural energy has become an inevitable choice. Wind power technology, as one of the most widely used and most mature new energy power generation technologies, has important significance for adjusting energy structure, meeting electricity demand, reducing environmental pollution and achieving sustainable development. However, after introducing wind power technology, the power grid presents the characteristics of complex structure, numerous branches, and variable operation mode, greatly increasing the risk of power grid failure. Therefore, when the power grid fails, timely and accurate fault diagnosis is crucial and has important significance for ensuring stable operation of the power grid and rapid recovery of power supply.

[0003] The power system frequency is an important operating parameter of the power system, which is closely related to the balance between active power supply and demand. When the power system fails, its frequency will change. With large-scale wind power integration, the wind turbine penetration rate increases. Due to the volatility of wind power, large-scale wind power integration will inevitably disrupt the original power supply and demand balance of the system. The frequency stability problem of the power system containing large-scale wind power is particularly prominent. Therefore, changes in frequency can be used to determine changes in the power grid with wind power integration, so as to timely adjust the wind power generation situation and effectively improve the power grid's ability to accept wind power.

[0004] At present, a variety of power grid fault diagnosis methods have been proposed at home and abroad. These methods are usually based on rules, neural networks, deep learning, data-driven techniques, etc. However, the above methods have significant shortcomings, such as the rule-based power grid fault diagnosis has deficiencies in actual engineering, poor learning ability, and poor model prediction effect; the data-driven power grid fault diagnosis model has poor interpretability, and in the case of massive data in large power grids, it is prone to cause dimension disaster, easily falls into local optimum in dealing with nonlinear problems, and requires longer response time. These shortcomings result in low efficiency and timeliness of power grid fault diagnosis, and insufficient ability to deal with complex faults. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present application is to provide a wind farm fault detection method based on frequency change. This method considers the different characteristics of frequency in time and frequency domains when different faults occur under different wind turbine penetration rates, has higher sensitivity and stronger anti-interference ability than other methods, and to some extent, eliminates the influence of noise and interference, improves the accuracy and reliability of fault detection.

[0006] To achieve the above object, the technical scheme adopted by the present application is:

[0007] A wind farm fault detection method based on frequency change, characterized in that it comprises the following steps:

[0008] Step A, extracting frequency time sequence features and frequency spectrum features under different wind turbine permeability, different charged load conditions and cutting conventional power supply conditions;

[0009] Step B, analyzing the dynamic change of frequency under different fault scenarios, and applying Fourier transform technology to explore the change characteristics of power spectrum density, specifically: by performing fast Fourier transform to convert time domain signal to frequency domain represented as F, as follows:

[0010]

[0011] In the above formula, f is the frequency domain axis, N is the number of sampling points, Fs is the sampling frequency, and Fs is 10000;

[0012] Then calculate the power spectrum P according to the following formula and draw the frequency image of frequency;

[0013]

[0014] Step C, real-time monitoring of the frequency of the power system, when the frequency changes, collecting the time sequence of the changed frequency and doing Fourier transform to get its power spectrum;

[0015] Step D, extracting and analyzing the features of the obtained time sequence and power spectrum, comparing to determine which type of fault is more likely to occur in the power system and making an alarm.

[0016] On the basis of the above scheme, the specific steps of step A are:

[0017] Step A1, extracting the size of the peak value in the frequency time sequence and the corresponding time t n , and obtaining the number of peak values n, and on this basis, finding the maximum value of the time sequence and the corresponding time t max , t min ;

[0018] Step A2, fitting the rising edge and falling edge of the obtained frequency time sequence into a straight line f=kt+b using the least square method, and calculating the slope k, wherein:

[0019]

[0020] In the above formula, t i is the time corresponding to the i-th data in the rising edge or falling edge section, f iis the frequency corresponding to the i-th data in the rising or falling edge section;

[0021] Step A3, set the standard deviation threshold to 0.01, use moving average and moving standard deviation to find the time when the frequency tends to be stable under different fan permeability, as shown in the following formula:

[0022]

[0023] In the above formula, size is the window size, M t is the moving average; t represents the time point, starting from size;

[0024] For a given same time series and window size size, moving standard deviation S t is:

[0025]

[0026] In the above formula, σ 2 is the variance;

[0027] Step A4, determine the time proportion of the frequency exceeding the fluctuation range under normal conditions, that is, compared with the reference value 50Hz, the time proportion of the frequency fluctuation greater than ±0.2Hz;

[0028] Step A5, use Fourier transform to obtain the power spectrum of the frequency under different fan permeability, and perform feature extraction analysis.

[0029] On the basis of the above scheme, the specific steps of step C are:

[0030] Step C1, fixed cycle sampling is performed on the frequency of the power system, and the approximate value of the signal zero-crossing time Δt is solved, as shown below:

[0031] Δt=-λT s

[0032] In the above formula, T s is the sampling time interval;

[0033] Where λ is:

[0034] f(λ)=a0λ 3 +a1λ 2 +a2λ+a3

[0035]

[0036] In the above formula, u k is the input signal u in the sampling point (t k , u k ) after the zero-crossing point;

[0037] Let f(lambda) = a0 lambda 3 +a1 lambda 2 +a2 lambda + a3 = 0, the Newton iteration method is used to find the root, and after two iterations, delta t is obtained;

[0038] Step C2, according to step C1, two adjacent zero-crossing time in the same direction delta t1 and delta t2 are calculated, and the current period T and the real-time frequency f of the power system are obtained from the following formula real , as follows:

[0039]

[0040] In the above formula, n represents the number of samples in a period, T s is the sampling interval;

[0041] Step C3, repeat the operation in step B, and obtain the power spectrum of the real-time frequency f through Fourier transform.

[0042] On the basis of the above scheme, the specific steps of step D are as follows:

[0043] Step D1, collect the time sequence of the frequency of the wind farm under normal working condition and under different fault conditions, and extract the time sequence characteristics and frequency domain characteristics of the frequency under these conditions, and normalize the time domain and frequency domain characteristics by using z-score method, specifically:

[0044]

[0045] In the above formula, x aver is the average value of each time domain and frequency domain characteristic; x i is each time domain and frequency domain characteristic value; N is the number of sampling points (total number of characteristic values);

[0046] The each time domain and frequency domain characteristic which needs to be normalized includes:

[0047] The frequency peak value and the time t n of occurrence, the peak value occurrence frequency n, the time t max and t min corresponding to the extreme value, the slope k corresponding to the rising edge and falling edge, the overshoot amount sigma p %, the adjustment time t s , the spectrum, the power spectrum density, the time corresponding to the steady image trend, etc.

[0048] The normalized value z is calculated according to the following two formulas:

[0049]

[0050] In the above two formulas, sigma t is the standard deviation;

[0051] Step D2 involves filtering the normalized time-domain and frequency-domain features obtained in step D1 using information gain and gain rate evaluators. Specifically:

[0052] The information gain estimator is:

[0053]

[0054] In the above formula, A is a sample set that can be divided into V classes and generate M branch nodes. Then, the sample contained in the Vth branch node is the sample in A that takes the value a on a. V The sample is denoted as A. V Ent(A) is the empirical entropy;

[0055] The gain rate estimator is:

[0056]

[0057] In the above formula, Gainratio(A,a) is the gain ratio, and IV(a) is the information gain at a specific point a.

[0058] Based on the information gain and gain ratio of each time domain and frequency domain feature, the top few scores in descending order of comprehensive score are selected to form a feature index set for subsequent model training.

[0059] The steps for obtaining the comprehensive score are as follows: the information gain and gain rate weights for each time domain and frequency domain are set to 0.4 and 0.6 respectively, and the scores are calculated accordingly.

[0060] Step D3 involves using the feature index set obtained in step D2 to train and test the classification model, specifically as follows:

[0061] The random forest algorithm is used to identify faults, and the output expression is as follows:

[0062]

[0063] In the above formula, L represents the number of decision trees in the random forest, and y i R(x) represents the output of each decision tree, while R(x) represents the output of the random forest.

[0064] The wind farm fault detection method based on frequency variation described in this invention has the following advantages:

[0065] This method comprehensively considers the different characteristics of the frequency in the time and frequency domains when different faults occur under different wind turbine penetration rates. Compared with other methods, it has the advantages of high sensitivity and strong anti-interference ability, and can eliminate the influence of noise and interference to a certain extent, thereby improving the accuracy and reliability of fault detection. Attached Figure Description

[0066] The present invention includes the following figures:

[0067] Figure 1 This is a flowchart illustrating the wind farm fault detection method based on power system power characteristics according to the present invention. Detailed Implementation

[0068] The present invention will be further described in detail below with reference to the accompanying drawings.

[0069] Step A. Extract frequency time series features and frequency spectrum features under different wind turbine penetration rates, different load conditions, and the condition of switching to conventional power supply.

[0070] A1. Extract the magnitude of the peak value and its corresponding time t from the time feature sequence of frequency. n The number of times the peak occurs, n, is obtained, and based on this, the maximum and minimum values ​​of the time series and their corresponding times, t, are found. m t n .

[0071] A2. Fit the rising and falling edges of the obtained frequency time series to a straight line using the least squares method, and calculate its slope k. Let the straight line fitted by a certain rising or falling edge be f = kt + b, where...

[0072]

[0073] A3. Use moving average and moving standard deviation to find the time when the frequency tends to stabilize under different wind turbine penetration rates, setting the standard deviation threshold to 0.01. For a given time series window t1, t2, t3…t… n And window size, moving average M t The formula for calculating time t is:

[0074]

[0075] For the same time series and window size k, the moving standard deviation S t The formula for calculating time t involves calculating the average value M of the data within the calculation window. t and variance σ 2 :

[0076]

[0077] Then, the moving standard deviation S t It is the square root of the variance:

[0078]

[0079] A4. Determine the time proportion of the frequency exceeding the fluctuation range under normal conditions, that is, the time proportion of the frequency fluctuation greater than ±0.2 Hz compared with the reference value 50 Hz. The standard of Power Quality Power System Frequency Allowable Deviation stipulates that the normal frequency of the power grid in China is 50 Hz, and for the power grid with a capacity of 3 million kilowatts and above, the deviation is not more than ±0.2 Hz. According to this provision, in the power grid with wind power generation introduced, the frequency fluctuation range allowed is 50±0.2 Hz.

[0080] A5. Use Fourier transform to obtain the power spectrum of the frequency under different wind turbine permeability for feature extraction analysis.

[0081] Step B. Analyze the dynamic changes of the frequency under different fault scenarios, and apply Fourier transform technology to explore the characteristics of the power spectrum density. By performing fast Fourier transform (FFT), the time domain signal is converted into frequency domain representation F, and the frequency domain axis is represented by f. The calculation method is as follows

[0082]

[0083] Where N is the number of sampling points, and Fs is the sampling frequency, which is 10000.

[0084] Then calculate the power spectrum P,

[0085]

[0086] Finally, draw the frequency image of the frequency.

[0087] Step C. Real-time monitoring of the frequency of the power system, when the frequency changes, collect the time series of the changed frequency and do Fourier transform to obtain its power spectrum. For the measurement of the frequency of the power system under dynamic conditions, high-precision frequency measurement algorithms under dynamic conditions can be used, mainly including algorithms for calculating the frequency and frequency change rate according to the signal period and algorithms for calculating the signal period according to the time width of the signal waveform at adjacent zero-crossing points.

[0088] C1. Use embedded devices for fixed period sampling, set the sampling time interval as T s , there are n sampling points between adjacent 2 zero-crossing points in the same direction Δt1, Δt2, then the period T of the power system and the real-time frequency f real of the power system are

[0089]

[0090] Since the true time Δt of the signal zero-crossing point cannot be directly obtained, it is assumed that the curve of the signal near the zero-crossing point is an approximate function P(x). In order to obtain higher accuracy, Newton's cubic interpolation polynomial is used to approximate the true value.

[0091] f(λ) = a0λ 3 +a1λ 2 +a2λ+a3 (8)

[0092] The coefficients are calculated as follows. For the input signal u(t), the four sampling points before and after the zero-crossing point are denoted as (t k-1 ,u k-1 ), (t k-2 ,u k-2 ), (t k-3 ,u k-3 ) before the zero-crossing point and (t k ,u k ) after the zero-crossing point. The coefficients can be solved using the four sampling points.

[0093]

[0094] The Newton cubic interpolation polynomial function is obtained as follows:

[0095]

[0096] wherein the relationship between λ and Δt is

[0097] Δt = -λT s

[0098] The coefficients are obtained as follows:

[0099]

[0100] The zero-crossing point is solved by setting f(λ) = a0λ 3 +a1λ 2 +a2λ+a3 = 0, and the root is solved by using the Newton iteration method,

[0101]

[0102] wherein the initial value λ0 can be obtained by the triangle similarity method,

[0103]

[0104] Substituting Δt = -λT s into the equation, we obtain

[0105]

[0106] After two iterations, Δt = -λT s is obtained.

[0107] C2. The two adjacent zero-crossing points in the same direction are calculated, and the current period T is obtained by substituting Δt1 and Δt2 into equation (15).

[0108] T = (n - 1)T s + Δt1+ (T s - Δt2) (15)C3. Finally, the real-time frequency of the power system can be obtained by f = 1 / T.

[0109] C4. Repeat the operation in step B, and the power spectrum of the real-time frequency can be obtained by Fourier transform.

[0110] Step D. Extract and analyze the features of the obtained time series and power spectrum, and determine which type of fault is more likely to occur in the power system and issue an alarm. The power system connected to wind power generation may have wind turbine faults, load fluctuations, conventional generator faults, and other types of faults. In order to accurately determine the type of fault, the frequency time series features and power spectrum features need to be classified according to the fault type and a database needs to be established. By comparing the time series features and power spectrum features of the real-time frequency with the data in the database, the type of fault can be determined.

[0111] D1. Collect the frequency time series of the wind farm under normal operation and under different fault conditions, and extract the time series features and frequency domain features of the frequency under these conditions. Normalize the time domain and frequency domain features. To eliminate the influence of dimension and numerical difference, z-score method is used for normalization. Take the time corresponding to the stable feature image as an example:

[0112] First, calculate the average value t aver as shown in equation (16).

[0113]

[0114] where N is the number of collected times.

[0115] Then calculate the standard deviation σ t as shown in equation (17).

[0116]

[0117] Finally, the normalized value z is obtained by equation (18).

[0118]

[0119] Normalize other features according to the above method.

[0120] D2. Feature Selection. The original dataset may contain irrelevant features, which can lead to redundancy in the diagnostic model structure and slow convergence. Therefore, feature selection is necessary to find the most valuable feature set from the original feature set. Information gain and gain ratio evaluators are used to filter these features.

[0121] Information gain represents the ability of a feature to classify data samples. Generally, the smaller the information gain of a feature, the lower the purity of the node partitioning using that feature, which is detrimental to the classification of data samples. Assume that the proportion of samples of class k in sample set A is p. k (k = 1, 2, ..., n), a certain feature a has M feature values, denoted as {a 1 a 2 ,…,a M Using this feature to classify the sample set A, it can be divided into M classes and generate M branch nodes. Then, the sample contained in the Mth branch node is the sample in A that takes the value a on a. M The sample is denoted as A. M The "information gain" after partitioning the sample set A using feature a is defined as the difference between the empirical entropy of A and the empirical entropy of A given feature a. The empirical entropy can be calculated using equation (19).

[0122]

[0123] Therefore, information gain can be expressed as

[0124]

[0125] Information gain criteria prioritize features with more eigenvalues. To avoid this, gain ratio is needed to correct for information gain. Gain ratio is defined as...

[0126]

[0127] Based on the information gain, gain ratio, and ranking of each feature in descending order of the overall score, the top 15 features are selected for subsequent model training.

[0128] D3. Finally, the classification model is trained and tested using the feature index set. The random forest algorithm is used to identify faults. The decision tree of the random forest randomly selects a certain number (assuming m = log2M) of features from the feature set of the node, and then selects the best feature from the subset for partitioning. Then, a voting strategy is used to summarize the results of multiple weak classifiers, and then the mode of its classification categories is taken as the final classification result. The output result expression is shown in Equation (22).

[0129]

[0130] wherein L is a tree of decision trees in the random forest, y i is the output result of each decision tree, and R(x) is the output result of the random forest.

[0131] The contents not described in detail in the present specification belong to the existing technology known to those skilled in the art.

Claims

1. A wind farm fault detection method based on frequency variation, characterized in that, Includes the following steps: Step A: Extract frequency time series features and frequency spectrum features under different wind turbine penetration rates, different electrical load conditions, and the condition of disconnecting conventional power supply; Step B involves analyzing the dynamic changes in frequency under different fault scenarios and applying Fourier transform techniques to explore the characteristics of power spectral density changes. Specifically, the time-domain signal is transformed into a frequency-domain representation F by performing a fast Fourier transform, as shown below: In the above formula, f is the frequency domain axis, N is the number of sampling points, Fs is the sampling frequency, and Fs takes the value of 10000; Then, calculate the power spectrum P according to the following formula and draw the frequency graph. Step C: Monitor the frequency of the power system in real time. When the frequency changes, collect the time series of the changed frequency and perform a Fourier transform to obtain its power spectrum. Step D involves extracting and analyzing the features of the obtained time series and power spectrum to determine which type of fault in the power system is more likely to occur and to issue an alert.

2. The wind farm fault detection method based on frequency variation as described in claim 1, characterized in that: The specific steps of step A are as follows: Step A1: Extract the magnitude of the peak value and the corresponding time t from the time feature sequence of the frequency. n The number of times the peak occurs, n, is obtained, and based on this, the maximum and minimum values ​​of the time series and their corresponding times, t, are found. max t min ; Step A2: Fit the rising and falling edges of the obtained frequency time series into a straight line f = kt + b using the least squares method, and calculate its slope k. in: In the above formula, t i f represents the time corresponding to the i-th data point in the rising or falling edge segment. i This represents the frequency corresponding to the i-th data point in the rising or falling edge segment. Step A3: Set the standard deviation threshold to 0.01, and use moving average and moving standard deviation to find the time when the frequency tends to stabilize under different wind turbine penetration rates, as shown in the following formula: In the above formula, size is the window size, M t This is a moving average; t represents a time point, starting from size; Given the same time series and window size, the moving standard deviation S t for: In the above formula, σ 2 The variance is calculated using the following formula: Step A4: Determine the proportion of time that the frequency exceeds the fluctuation range under normal conditions, that is, the proportion of time that the frequency fluctuation is greater than ±0.2Hz compared with the reference value of 50Hz. Step A5: Use Fourier transform to obtain the power spectrum of the frequency under different wind turbine penetration rates, and perform feature extraction analysis.

3. The wind farm fault detection method based on frequency variation as described in claim 1, characterized in that: The specific steps of step C are as follows: Step C1 involves sampling the power system frequency at a fixed period and calculating the approximate value Δt for the signal's zero-crossing time, as shown below: Δt=-λT s In the above formula, T s The sampling time interval; Where λ is represented as: f(λ)=a0λ 3 +a1λ 2 +a2λ+a3 In the above formula, u k The sampling points (t) after the zero point k ,u k The input signal u in ); Let f(λ) = a0λ 3 +a1λ 2 +a2λ+a3=0, the root is found by Newton's iteration method, and Δt is obtained after two iterations; Step C2: Calculate the two adjacent zero-crossing times Δt1 and Δt2 in the same direction as in step C1, and then obtain the current period T and the real-time frequency f of the power system using the following formula. real As shown below: In the above formula, n represents the number of samples in the period, and T s The sampling interval; Step C3: Repeat the operation in step B, and obtain the real-time frequency f through Fourier transform. real The power spectrum.

4. The wind farm fault detection method based on frequency variation as described in claim 1, characterized in that: The specific steps of step D are as follows: Step D1 involves collecting time series data of wind farm frequencies during normal operation and under different fault conditions, extracting the temporal and frequency domain features of these frequencies, and normalizing the time and frequency domain features using the z-score method. Specifically: In the above formula, x aver x represents the average value of each time-domain and frequency-domain characteristic; i For each time-domain and frequency-domain feature value; N is the number of sampling points; The standardized value z is obtained by calculating using the following two formulas: In the above two equations, σ t Standard deviation; Step D2 involves filtering the normalized time-domain and frequency-domain features obtained in step D1 using information gain and gain rate evaluators. Specifically: The information gain estimator is: In the above formula, A is a sample set, divided into M classes and generating M branch nodes. Then, the sample contained in the Vth branch node is the sample in A that takes the value a on a certain feature a. V The sample is denoted as A. V Ent(A) is the empirical entropy; The gain rate estimator is: In the above formula, Gainratio(A,a) is the gain ratio, and IV(a) is the information gain at feature a; Based on the information gain and gain ratio of each time domain and frequency domain feature, the top few scores in descending order of comprehensive score are selected to form a feature index set for subsequent model training. The steps for obtaining the comprehensive score are as follows: the information gain and gain rate weights for each time domain and frequency domain are set to 0.4 and 0.6 respectively, and the scores are calculated accordingly. Step D3 involves using the feature index set obtained in step D2 to train and test the classification model, specifically as follows: The random forest algorithm is used to identify faults, and the output expression is as follows: In the above formula, L represents the number of decision trees in the random forest, and y i R(x) represents the output of each decision tree, while R(x) represents the output of the random forest.

Images