Wind speed sensing abnormal data recognition method for wind turbines based on speed correlation constraints

Abstract

The present invention relates to a method for identifying abnormal data of wind speed perception of a wind turbine based on speed association constraints, comprising the following steps: 1) obtaining wind speed sequence data of a target unit; 2) using a binary morphology segmentation algorithm to detect the time series change point of the wind speed sequence, and using the time series change point as a segmentation point to divide the wind speed sequence into multiple subsequences in time series; 3) performing abnormal data identification on the divided subsequence data based on speed and boundary constraints. Compared with the prior art, the present invention has the advantages of taking into account both economy and feasibility, accurate and fast evaluation, etc.

Description

Technical Field

[0001] The invention relates to the field of wind speed monitoring of wind turbines in wind farms, and in particular to a method for identifying abnormal data of wind speed perception of wind turbines based on speed association constraints. Background Art

[0002] At present, a lot of research has been done on wind farm data cleaning and many results have been achieved. It is summarized that most of the existing wind farm data cleaning methods are centered around the characteristics of the wind speed-power curve. Data cleaning is carried out by analyzing the spatial location and morphological characteristics of different types of abnormal data distributed on the wind power curve, such as: quartile-clustering method, confidence equivalent boundary model, optimal within-group variance method, change point-quartile method, Thompson tau-quartile method, local particle swarm algorithm, etc. The above existing methods can effectively identify the accumulation type abnormal power data that obviously violates the wind speed-power characteristics outside the power curve, solve the problem of wind power data accuracy and effectiveness in the process of modeling and analyzing the operating characteristics of wind turbines, and are of great significance to the evaluation of the power generation performance of the unit.

[0003] However, considering that the uncertainty in the wind speed-power conversion relationship is different from the mechanism of wind speed uncertainty, the data cleaning constraint boundary constructed by the wind speed-power correlation characteristic is much larger than the constraint boundary of normal wind speed data. For example, in the wind power curve of a certain wind turbine, the normal wind speed range corresponding to the point with a power value of 1000kW is 5.5-9.1m / s, and the interval width is 3.6m / s. For wind speed data, if the deviation of a single point measurement value reaches 3.6m / s, it has been defined as abnormal data. Given that the variable range of data within the normal boundary and the sensitivity to abnormal data are different for both, the constraint conditions constructed by the wind speed-power correlation characteristic with a large variable range and small sensitivity to clean the wind speed data with high sensitivity and small variable range are likely to cause the weak fault wind speed data generated by the performance degradation of the wind speed sensor located inside the wind power curve to be mistaken as normal data, affecting the subsequent application of the wind speed data. Therefore, it is necessary to propose a deep cleaning method for wind speed data based on the distribution characteristics of the wind speed data itself, which has high value for improving the quality of wind speed measurement data in wind farms and comprehensively evaluating the operating performance of wind speed sensors. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method for identifying abnormal wind speed perception data of a wind turbine generator set based on speed correlation constraints.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for identifying abnormal data of wind speed perception of a wind turbine generator set based on speed association constraints comprises the following steps:

[0007] 1) Obtain wind speed sequence data of the target unit;

[0008] 2) The time series change points of the wind speed sequence are detected by using a binary morphological segmentation algorithm, and the time series change points are used as segmentation points to divide the wind speed sequence into multiple subsequences in time series;

[0009] 3) Identify abnormal data on the divided subsequence data based on speed and boundary constraints.

[0010] The abnormal data types include sudden abnormal wind speed data and gradual abnormal wind speed data.

[0011] The generation situations of the mutation type abnormal wind speed data include:

[0012] (1) Since the wind speed sensor's ability to resist external electromagnetic disturbances is weakened, when the external disturbance is strong, the measured data will undergo a large-scale mutation, the single abnormality will last for a short time, and the distribution will be discrete;

[0013] (2) Due to the damage of important parts of the wind speed sensor, the abnormal state has a non-trend mutation, the measured wind speed has a sudden change, the distribution is relatively continuous, and the number is large;

[0014] The generation of the progressive abnormal wind speed data includes:

[0015] (1) The degree of damage to the wind speed sensor components is a gradual process, and the measured data deviates from the normal value. The offset gradually increases with the evolution of the abnormal state of the equipment, that is, weak fault data;

[0016] (2) Consecutive identical values ​​due to communication errors.

[0017] In the step 2), determining the timing change point specifically includes the following steps:

[0018] 21) Calculate the local statistics of the wind speed sequence, and determine the final change point based on the three constraints of satisfying the local statistic maximum, local peak point and statistic bound;

[0019] 22) dividing the entire segment of the wind speed sequence data into two segments according to the final change point, and repeating step 21) for each segment to perform segmentation again;

[0020] 23) Repeat steps 21) and 22) until no new change points appear.

[0021] The step 21) specifically comprises the following steps:

[0022] 211) For the local data of wind speed sequence V = {V S ,V S+1 ,…,VT}, calculate the corresponding local statistics T h k , k=h+1,…,Th, obtain the local statistics T h k The point k0 corresponding to the maximum * , and mark this point as a suspected change point;

[0023] 212) If the suspected change point meets the determination condition of the local peak point, the suspected change point is determined to be a local peak point. The determination condition of the local peak point is specifically:

[0024] Taking the suspected change point as the center, the points within the window width of 2h are differentiated. If the sum of the differences within the range of h on the left of the suspected change point that are greater than 0 and the differences within the range of h on the right that are less than 0 accounts for the proportion of the window width 2h (k0 * ) is greater than the threshold γ, then the suspected change point is judged to be a local peak point;

[0025] 213) Define the boundary that distinguishes change points from non-change points a , and according to the boundary a The final changing points are obtained after removing the invariant points from the local peak points.

[0026] In step 211), the local statistic T h k The calculation formula is:

[0027]

[0028]

[0029]

[0030] Among them, C represents the search coefficient, h represents the search window width, which is related to the sample size T, k represents the point to be detected, and v i 、v j are the i-th and j-th points in the local data respectively.

[0031] In the step 212), the specific gravity P (k0 * ) is:

[0032]

[0033]

[0034]

[0035] Among them, df(i,j) represents the statistics of two points i and j and The difference of df(i,j) = 1, 0, and -1 respectively indicate that the statistic is increasing, stable, and decreasing. df(i,j)＝1 Indicates that the point is the count of the rising point of the statistic, I df(i,j)＝-1 Indicates that the point is the count of the statistical drop point, and the threshold γ is 0.7 or 0.8.

[0036] In step 213), the boundary ξ a Defined as:

[0037]

[0038] Among them, a is a parameter used to distinguish non-changing points.

[0039] In the step 3), the sudden abnormal wind speed data is identified by combining the multi-interval speed constraints and the wind speed upper and lower limit boundary constraints, and the data that does not satisfy the wind data speed change constraint and the wind speed upper and lower limit boundary constraints are determined as sudden abnormal wind speed data. The wind speed upper and lower limit boundary constraints are specifically:

[0040] v qmin ≤v i ≤v qmax

[0041] Among them, v qmin and v qmax represent the upper and lower quartiles of wind speed, respectively;

[0042] The multi-interval speed constraint S is a set of speed constraint intervals. For the rth time window, the speed constraint boundary of the wind speed sequence is determined as follows:

[0043]

[0044] Among them, ε u , ε d are sensitivity coefficients, v min and v max Indicates the minimum and maximum wind speeds of wind farm statistics, are the maximum and minimum values ​​of the ascent speed constraint, are the maximum and minimum values ​​of the descent speed constraint respectively.

[0045] In the step 3), the progressive abnormal wind speed data is identified based on multi-machine association matching, specifically:

[0046] A wind speed mutual diagnosis community is formed by multiple adjacent units in the same wind belt, and a wind speed sequence correlation description model based on the Copula function is constructed. Multiple units adjacent to the target unit are selected as associated seed units, and the correlation between the target unit and the seed unit group in different wind speed subsequences is calculated one by one. If the correlation between the target unit and the seed unit group in a wind speed subsequence is lower than 0.5 and the correlation between the seed units is higher than 0.75, the wind speed subsequence corresponding to the target unit is judged to be progressive abnormal data. The calculation formula of the correlation coefficient is:

[0047]

[0048] Among them, P(V1, V2) is the correlation coefficient of the wind speed subsequences V1 and V2 of the target unit and the seed unit, and C(u, v) is the joint probability distribution function.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] 1. The present invention adopts a change point detection method based on binary morphological segmentation to divide the global wind speed sequence into several independent wind speed subsequences, and takes short-time scale wind speed data as the research object of anomaly analysis, which can improve the accuracy and reliability of the data cleaning method.

[0051] Second, the data cleaning method based on speed constraints and wind speed spatiotemporal correlation constraints of the present invention can accurately identify abnormal perceived wind speeds of various categories in wind farms, especially weak fault abnormal wind speeds with less obvious abnormal characteristics;

[0052] 3. The data cleaning method of the present invention is based on the spatiotemporal correlation characteristics of wind speed. It can quickly and accurately complete the wind speed data cleaning work on the basis of retaining the original normal data, which is of great significance for comprehensively improving the quality of wind speed data and quantitatively evaluating the operating performance of sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the distribution characteristics of abnormal wind speed data from the sensor.

[0054] Figure 2 Figure (2a) is the wind speed data correlation distribution characteristics, where Figure (2b) is the wind speed data curve and Figure (2b) is the wind speed correlation coefficient.

[0055] Figure 3 It is a schematic diagram of change point detection, wherein Figure (3a) is a step flowchart and Figure (3b) is a schematic diagram of data partitioning.

[0056] Figure 4 The framework of the abnormal wind speed identification method.

[0057] Figure 5This is an example of a multi-interval speed constraint.

[0058] Figure 6 For anomaly data detection based on velocity and boundary constraints.

[0059] Figure 7 This is the flow chart for cleaning abnormal wind speed data.

[0060] Figure 8 This is the wind speed sensor status diagnosis flowchart.

[0061] Fig. 9 T is the wind speed series h k Statistics.

[0062] Fig.10 Partition results for wind speed series.

[0063] Fig.11 are the mean and variance statistics of wind speed in each series.

[0064] Fig.12 is the wind speed correlation statistics of the wind speed subsequence. DETAILED DESCRIPTION

[0065] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] Example

[0067] The present invention provides a method for identifying abnormal wind speed perception data of a wind turbine based on speed correlation constraints. Based on the temporal and spatial correlation characteristics of wind speed data, a method and architecture for cleaning abnormal wind speed data of a wind farm are proposed, and weak fault wind speed data with less obvious abnormal characteristics is deeply mined, aiming to improve the overall quality of wind speed data measured in the wind farm, and at the same time provide a reference for the operating performance assessment of wind speed sensors in different speed ranges.

[0068] This method first summarizes and classifies the common abnormal wind speed data of wind farms. Considering that long-time scale wind speed data is prone to cause coverage and annihilation of some local feature information in the process of feature analysis and association modeling, the change point detection algorithm based on binary morphological segmentation is used to divide the global wind speed sequence into time series intervals. For each subsequence, an abnormal wind speed recognition method based on speed-association constraints is constructed from the time and space dimensions to carry out data cleaning. Finally, the actual measurement data of the wind farm is used as the research object to verify the effectiveness of the method. The specific technical content is introduced as follows:

[0069] 1. Wind speed data cleaning

[0070] 1.1 Characteristics and classification of abnormal wind speed data

[0071] The wind speed data of wind turbines is mainly captured by wind speed sensors exposed on the outside of the fuselage. The operating status and performance quality of wind speed sensors directly affect the quality of the measured data. In actual wind farms, lightning strikes, bird strikes, material corrosion, mechanical wear, etc. can cause sensor performance degradation, resulting in inaccurate rotation speed or even stop rotation. According to the trend of data changes, typical abnormal wind speeds can be classified into two categories, as shown in Table 1:

[0072] Table 1 Classification and causes of abnormal wind speed in wind farms

[0073]

[0074] (1) Mutation-type abnormal data: There are two main situations in which this type of abnormal data is generated. One is that the wind speed sensor's ability to resist external electromagnetic disturbances is weakened. When the external disturbance is strong, it is easy to cause a large-scale mutation in the measured data. The duration of a single abnormality is often short and the distribution is discrete. The other situation is that the important components of the wind speed sensor are damaged, resulting in a non-trend mutation of the abnormal state. The measured wind speed mutates, the distribution is relatively continuous, and the number is large. For example, the transmission bearing of the wind speed sensor is broken and the sensor has no input data.

[0075] (2) Progressive abnormal data: There are also two types of abnormal data. One is that the degree of damage to the wind speed sensor components is a gradual process, and the measured data deviates from the normal value. The offset gradually increases with the evolution of the abnormal state of the equipment, that is, weak fault data. For example, the performance of the transmission bearing lubricant deteriorates, causing the sensor transmission to become stuck and inflexible, resulting in data inaccuracy. The other is continuous identical values ​​caused by communication errors. The generation of such abnormalities is random.

[0076] In actual wind farms, the wind turbine SCADA system is equipped with a sensor fault diagnosis module. The basic logic of diagnosis is to set upper and lower wind speed thresholds. When the sensor measurement value exceeds the threshold range, the system alarms and the unit can identify abnormal data of the sensor exceeding the threshold. For progressive abnormal data, the range of wind speed data deviation from the normal wind speed is related to the evolution process of the abnormal state of the sensor. The abnormal wind speed is still within the threshold range of the system diagnosis, and the SCADA system cannot accurately identify such abnormal wind speed.

[0077] 1.2 Introduction to wind speed data cleaning method with speed-correlation constraints

[0078] Theory shows that the time-varying characteristics of wind speed are actually the externalization of complex atmospheric movement. The wind speed at a certain location in space at a certain moment has a certain correlation with its historical state and the wind speed at adjacent locations. The shorter the time difference and the closer the spatial distance, the more obvious the correlation of wind speeds is. For multiple wind turbines that are adjacent in space and located in the same wind belt, their wind measurement sequences show a strong wind speed correlation, especially when the monsoon climate is more obvious, the wind speed correlation is higher. Assuming that there are several units in the same wind belt, the wind speed sensor of one of the units is abnormal, and the measured wind speed deviates from the actual wind speed. The spatial correlation between the measured wind speed data of the abnormal equipment and the measured wind speed data of other normal equipment in the surrounding area will be significantly reduced, while the wind speed correlation between other normally operating units will not be affected and will remain at the original level. In theory, it is feasible to identify abnormal wind speed by analyzing the spatiotemporal correlation of wind measurement data of multiple adjacent wind speed sensors.

[0079] Analyzing the actual measured wind speed data, a certain unit cluster consists of three units located in the same wind belt and adjacent to each other. The units are numbered #1, #2 and #3. Figure (2a) shows the measured wind speed curve of the cluster. At the 123rd time point, the wind speed sensor of unit #3 has an abnormality. The wind speed sequence is divided into two subsequences [0,122] and [123,144] using the abnormal point as the time series segmentation point. The global wind speed correlation coefficient and local wind speed correlation coefficient among the three units are calculated. The calculation results are shown in Figure (2b). It can be seen that the global wind speed correlation between units ([0,144]) and the local wind speed correlation in the normal time series interval ([0,122]) are both above 0.75, indicating a high correlation. In the abnormal time series interval ([123,144]), the wind speed correlation between unit #3 and units #1 and #2 is significantly reduced, indicating an irrelevance. The wind speed correlation between units #1 and #2 is still highly correlated.

[0080] It can be seen that the generation of abnormal wind speed data will cause the wind speed correlation between units to decrease, especially in the abnormal interval, the local wind speed correlation decreases significantly, characterized by irrelevance, which fully proves the feasibility of abnormal wind speed identification based on wind speed correlation theory. In addition, although the global wind speed correlation between units has also decreased to a certain extent compared with the correlation in the normal interval, it is still characterized by high correlation. This is because using long-time scale wind speed data as the object of correlation modeling and analysis is likely to cover some local feature information, thereby ignoring some short-time scale abnormal feature quantities. Therefore, before constructing a data cleaning algorithm based on wind speed spatiotemporal correlation, it is necessary to first divide the global wind speed data into time series intervals, divide the global quantity into several local subsequences, and then use each local wind speed subsequence as the object of data cleaning modeling and analysis to improve the accuracy of the algorithm.

[0081] 2. Time series interval division of wind speed data

[0082] As mentioned above, compared with normal wind speed data, the mean, variance and other multivariate statistics of abnormal wind speed data will change significantly. The time series division problem of the global wind speed sequence can be transformed into a multi-change point detection problem. The time series points where the multivariate statistics of the wind speed sequence change significantly are detected as change points. With the change points as the time series segmentation points, the global variables can be divided into several sections of independent local variables. Based on this, the present invention uses a binary morphological segmentation algorithm to detect the time series change points of the wind speed sequence, and uses the change points as the time series points for interval division to complete the interval division of the global wind speed sequence.

[0083] The problem of statistical analysis of change points originated from the problem of quality control in management science. Now it is widely used in many fields such as economics, finance, and engineering. The change point is the moment when the random nature of a process changes. Assume a set of independent wind speed sequences V0, V1, …, V i ,…,V T , if k(0≤k≤T) satisfies: V i ~F0(i=0,1,2,…,k),V i ~F1(i=k+1,k+2,…,T), where F0 and F1 are two different distribution functions, then k is called the change point.

[0084] In multiple change point detection, the binary segmentation (BS) algorithm is one of the most classic algorithms, and the algorithm has high stability. However, since the BS algorithm uses the global CUSUM statistic, when the time series is long, it is difficult to effectively identify the local change point problem. In view of the stability advantage of the BS method, the present invention is based on the BS sequential idea and uses local statistics as the basic statistics. Secondly, by analyzing the morphology of the local statistics, it is found that the statistic at the change point is not only the local maximum value, but also the local peak value. Therefore, the present invention adds a constraint condition for change point detection, namely, the binary morphological segmentation condition:

[0085] (1) The local statistic is the largest;

[0086] (2) local peaks;

[0087] (3) The statistic satisfies the bounds;

[0088] Only when the above three conditions are met at the same time can it be considered a change point, thereby increasing the robustness of change point identification.

[0089] Assume that the wind speed sequence V collected by a wind speed sensor is V = {V0, V1, ..., V T}, the description model of the sequence is defined as follows:

[0090] V i=μ i +ε i ,i=0,1,2,...,T (1)

[0091] Among them, V i is the wind speed sequence of the ith time subinterval, and its corresponding mean is μ i , assuming μ0=μ1=μ2=…=μ q ≠μ q+1 =…=μ w ≠μ w+1 =…=μ e ≠μ e+1 =…=μ n , the number and position of the change points in the sequence are unknown. In practice, the number of change points N is much smaller than the sample size T, that is, N = o(T). In order to balance the efficiency and accuracy of the algorithm, the present invention adopts a change point binary segmentation detection algorithm based on morphological recognition. The idea of ​​the algorithm is still based on the BS algorithm. In the process of change point detection, the local statistic T is first calculated. h k When the value of a point statistic is greater than the set threshold, the point is taken as the first suspected change point. Assume that the first suspected change point is k0 * , the time interval [0,T] is k0 * The treatment is divided into two sub-segments {V0, V1, ..., V k0*} and {V k0* ,…,V T}; Then continue to calculate the local statistics in the two newly divided intervals to detect suspected change points, and then perform binary division based on the found suspected change points, and repeat the cycle recursively until all points are found.

[0092] The specific steps of the change point binary segmentation detection algorithm based on morphological recognition are as follows: Figure 3 As shown, it is mainly divided into:

[0093] (1) Detect and determine suspected change points

[0094] a) For local data V = {V S ,V S+1 ,…,V T}, first calculate the local statistics T h k , k=h+1,…,Th, find the h k The largest point k0 * , mark this point as a suspected change point, T h k and k0 * The calculation of is as follows:

[0095]

[0096]

[0097]

[0098] Wherein, C represents the search coefficient, and the optional range is [0.3, 0.5]. In this example, C = 0.4; h represents the search window width, which is related to the sample size T; k represents the point to be detected.

[0099] b) To verify whether the suspected change point is a change point, the point must also meet the peak judgment conditions. The peak judgment method is to take this point as the center and make differences in the window width of 2h. If the sum of the differences within the range of h on the left of the suspected change point that is greater than 0 and the differences within the range of h on the right that is less than 0 accounts for the proportion of the window width 2h PR (k0 * ) is greater than γ, then the point is considered to be a peak, and the peak identification statistic is defined as follows:

[0100]

[0101]

[0102]

[0103] Where df(i,j) represents T h i and The difference of the two-point statistic, df(i,j) = 1, 0, -1 respectively indicates that the statistic is increasing, stable, and decreasing; I df(i,j)＝1 Indicates that the point is the count of the rising point of the statistic, I df(i,j) =-1 indicates that the point is the count of the statistical decrease point; P(k0 * ) represents the weight of the point as a peak. If P(k0 * )>γ, then k0 * It is a single peak. Due to the randomness of the data and the distribution form of the data statistics, γ is generally set to 0.7 or 0.8 in practice. Since the local statistics distribution has an obvious peak at the change point, this judgment method is not likely to miss the real change point;

[0104] c) The points selected by the first two steps are all peak points that meet the local maximum value, including the change point. Since the statistical values ​​of the selected points at the change point and the non-change point are quite different, it is only necessary to roughly define a boundary ξ to distinguish the change point from the non-change point. a , the final change point can be determined. aIt is related to the statistical value. The selection criterion is to filter out non-changing points. Since the test statistic has a peak at the changing point, the non-changing points around the changing point do not constitute a local maximum. Therefore, it is sufficient to remove the non-changing points far away from the changing point. The statistical value of the non-changing points far away from the changing point is much smaller than that at the changing point. Therefore, for the boundary ξ a The definition is as follows:

[0105]

[0106] The range of a is very wide, and it is not sensitive to the accuracy of the result to a certain extent. In this example, a=0.4 is set. If k0 * If the above three conditions are met, then k0 * For the change point.

[0107] (2) Data segmentation

[0108] The whole segment of data V={V0,V1,…,V i ,…,V T}With the change point k0 * Divided into two segments {V0, V1, ..., V k0*} and {V k0* ,…,V T}, and then execute the algorithm of step (1) on these two segments respectively.

[0109] (3) Repeat (1) and (2) until no new change points appear.

[0110] The change point detection algorithm based on binary morphological segmentation can divide the global wind speed sequence into multiple wind speed subsequences with different distributions. Taking the wind speed subsequences as the research object can make the local characteristics of the wind speed data more obvious, and can completely and accurately identify the similarities and differences between the data. In addition, using the distribution characteristics of the wind speed data as the basis for interval division, the normal wind speed sequence and the abnormal wind speed sequence can be separated in time series. Figure 2 Taking the data as an example, the change point positions calculated by the proposed change point detection method are the 47th, 63rd and 123th time series points, which are basically consistent with the set abnormal points. The 47th and 63rd time series points may be due to the obvious time-varying characteristics of wind speed caused by meteorological changes, which makes the multivariate statistics of the two wind speed sequences change significantly. Therefore, after the global wind speed is divided by the change point detection algorithm based on binary morphological segmentation, each subsequence has the following three states:

[0111] (1) Normal data;

[0112] (2) Normal data is mixed with a small amount of abnormal data of mutation type;

[0113] (3) Abnormal data.

[0114] 3. Theoretical framework of wind speed data cleaning method

[0115] As mentioned above, abnormal wind speed data can be divided into mutation abnormal data and progressive abnormal data according to their distribution characteristics. For mutation abnormal data and some progressive abnormal data (continuous identical values), the present invention constructs an abnormal data recognition algorithm based on speed constraints based on the time correlation characteristics of wind speed data. In addition, in order to avoid misdiagnosis of gusty climate, the present invention adds wind speed boundary threshold constraints; for progressive abnormal data such as wind speed inaccuracy and instability, the present invention constructs correlation constraints based on the spatial correlation of wind speed, uses Copula functions to describe the spatial correlation of wind speed, and identifies such abnormal data by checking the consistency of wind speed correlation of multiple units. The architecture of the abnormal data recognition method is as follows: Figure 4 shown.

[0116] 3.1 Principles of data cleaning methods based on speed and boundary constraints

[0117] Wind speed data has autocorrelation on a time scale. The current data is associated with its historical state. If the data changes suddenly over a large range due to electromagnetic interference, communication failure, etc. at a certain moment, it contradicts the autocorrelation. Therefore, the present invention constructs a wind speed data change speed constraint condition to identify such abnormal data. The model is described as follows:

[0118] Assume that the wind speed sequence V = {v i ,v i+1 ,…,v n},v i Indicates the tth i Wind speed at the moment, speed constraint s of the sequence r =[s r min ,s r max ] represents the upper and lower thresholds of the wind speed data change rate, where s r min Refers to the minimum speed of wind data change in this constraint, s r max Refers to the maximum speed of wind data change in this constraint, using s r min Find consecutive identical outliers, s r max Look for mutation outliers. The wind data that meets the speed constraint is actually consistent with the temporal correlation of wind speed. The speed of change of wind speed s(v i ,v j ) and the constraints on data normality are as follows:

[0119]

[0120]

[0121] Take the actual data collected by the wind speed sensor of a real wind farm as an example. Figure 5 As shown, given the speed constraint interval s 1up =[s 1up min ,s 1up max ] and s 1down =[s 1down min ,s 1down max ], the change speed of the data point pairs s(v0,v1),s(v1,v2),…,s(v8,v9) satisfies the speed constraint interval s 1up =[s 1up min ,s 1up max ]. Similarly, the data point pair s(v9,v 10 ),s(v 10 ,v 11 ),…,s(v 12 ,v 13 ) satisfies the speed constraint interval s 1down =[s 1down min ,s 1down max ], therefore, the above data is judged as normal data, and the data point pair s(v 13 ,v 14 ) does not satisfy the constraint interval s 1down =[s 1down min ,s 1down max ], so we determine the data point v 14 is abnormal data.

[0122] A multi-interval speed constraint S refers to a set of speed constraint intervals s r (r=1,2,…,m), that is, S={s1,s2,…,s m}. In a given time window r, the data point (v i ,v j ) satisfies the multi-interval velocity constraint S: any data point in the window (v i ,v j ) satisfies the speed constraint s r The multi-interval speed constraint can be used to identify communication errors and electromagnetic interference in the wind speed sensor. The abnormal data is manifested in data mutation and continuous identical values. For the rth time window, the wind data V rThe velocity constraint boundary is determined as follows:

[0123]

[0124] In the formula, ε represents the sensitivity coefficient. In this example, ε is taken as u =0.5; ε d =0.01; v min and v max Indicates the minimum and maximum wind speeds of wind farm statistics.

[0125] Considering that wind speeds can also vary widely in gusty climates, e.g. Figure 6 As shown in the figure, the wind speed curves of the three units over a period of time are shown. It can be seen that at the 47th time point, the wind speeds of the three units all changed greatly, which was caused by gusts; at the 67th time point, only the wind speed data of unit #3 changed suddenly, which was affected by electromagnetic interference. Therefore, only using speed constraint detection will cause misdiagnosis of gusty climate.

[0126] Analysis of the distribution characteristics of mutation abnormal data shows that such abnormal data often deviates from normal values ​​in a large range. Based on this, the present invention adds upper and lower limit constraints on wind speed:

[0127] v qmin ≤v i ≤v qmax (12)

[0128] In the formula, v qmin and v qmax represent the upper quartile and lower quartile of wind speed respectively.

[0129] The wind data that does not meet the wind data speed change constraint such as formula (10) and the upper and lower bound constraints such as formula (12) are judged as sudden abnormal wind speed. Figure 6 As shown, based on speed constraints and boundary constraints, misdiagnosis can be avoided and the accuracy and universality of the algorithm can be improved.

[0130] 3.2 Principle of data cleaning method based on association constraints

[0131] For progressive abnormal wind speed data, its distribution is continuous, the number is relatively large, and the distribution characteristics are significantly different from the actual wind speed. For a unit community composed of multiple adjacent wind turbines in the same wind belt, the wind speed correlation between the abnormal sensor and other normal sensors will be significantly reduced after the abnormality occurs. In addition, it is a low-probability event that the wind speed sensors of multiple adjacent units are abnormal at the same time. Therefore, considering the practicality and accuracy of the algorithm, the present invention takes multiple adjacent units as the object to constitute an abnormal wind speed mutual diagnosis community, constructs a wind speed sequence correlation description model based on the Copula function, selects multiple units adjacent to the target unit as associated seeds, and identifies the abnormal wind speed by analyzing the correlation between the target and the seed units.

[0132] Copula function theory was originally applied to the correlation analysis of complex random variables in the financial industry. Due to the strong practicality and high accuracy of the algorithm, it has been widely used in the correlation analysis of wind speed in wind farms in recent years. The theory is as follows: Assume a distribution function H(·) that describes a random process. The random process has n random variables x1, x2, …, x n , whose marginal probability distribution functions are F1(x1),…,F n (x n ), then there exists a Copula function C(·) that satisfies:

[0133] H(x1,x2,...,x n )=C(F1(x1),F2(x2),...,F n (x n )) (13)

[0134] If the marginal probability distribution function F1(x1),…,F n (x n ) is continuous, then C(·) is uniquely determined, and vice versa. Among them, C(·) describes the distribution of correlation between multiple variables from the perspective of probability, which is suitable for the distribution of any random variable. If a monotonically increasing transformation is made to the variable, the corresponding C(·) will not change, so the values ​​of the consistency and correlation measures obtained from C(·) will not change. Compared with the Pearson coefficient, the correlation measure derived from the Copula function has a wider range of applications and is more practical.

[0135] Commonly used Copula functions include multivariate normal Copula function, Archimedean Copula function and multivariate t-Copula function. Among them, the multivariate t-Copula function can be used to calculate the correlation characteristics between wind speed sequences, and can also express the symmetrical upper tail or lower tail correlation of wind speed. Therefore, this example uses the t-Copula function to describe the correlation of wind speed sequences. The calculation process is as follows:

[0136] Assuming that the wind speed sequence (V1, V2) measured by the two wind speed sensors obeys the Weibull distribution, its distribution function is:

[0137]

[0138] Where V is the wind speed; c and k are the scale parameter and shape parameter of the Weibull distribution, respectively.

[0139] Assume u = F1(x), v = F2(x). Since the wind speed sequence is continuous, there exists a t-Copula function C(u,v) that connects the joint distribution and marginal distribution of the two wind speed sequences. The calculation formula of the joint probability distribution function is:

[0140]

[0141] Where λ and ρ are the parameter estimates of the Copula function.

[0142] The joint distribution function C(u,v) is obtained, and the correlation between different wind speed sequences can be quantitatively evaluated by the Kendall rank correlation coefficient. The correlation coefficient calculation formula is:

[0143]

[0144] The strength of the correlation between the two units can be further determined based on the calculated correlation coefficient.

[0145] The unit corresponding to the wind measurement sequence to be detected is selected as the target unit, and the unit adjacent to the target unit and in the same wind belt is the associated seed unit group. Based on the constructed correlation model, the correlation between the target unit and the seed unit group under different wind speed subsequences is calculated segment by segment. If the correlation between the target unit and the seed unit group under a certain wind speed subsequence is low while the correlation between the seed units is high, the wind speed subsequence corresponding to the target unit is judged to be abnormal data.

[0146] 4. Wind speed data cleaning process

[0147] The cleaning process of abnormal wind speed data is mainly divided into three parts: wind speed data time series interval division, mutation type abnormal data identification and progressive data state identification.

[0148] (1) Wind speed data time series interval division

[0149] The distribution characteristics of wind speed data are identified and segmented using the change point binary detection technology based on morphological features. The wind speed sequence is divided into several wind speed subsequences according to its distribution characteristics. For the measured wind speed V of unit j, ij ] p*T , v ijrepresents the measured wind speed of unit j at the i-th moment, and the divided wind speed sequence can be expressed as V j = {V j1 ,V j2 ,…,V jn},V jn represents the nth subsequence of unit j.

[0150] (2) Identification of mutation-type abnormal data

[0151] As mentioned above, the mutation type abnormal state appears to be highly random and has a large range of mutations. The present invention adopts a detection technology based on data change speed and boundary constraints. The algorithm flow includes: determining speed constraints, determining upper and lower boundary constraints of data, calculating wind speed data differences, sliding detection, and abnormal wind speed reconstruction based on interpolation algorithm.

[0152] (3) Identification of progressive abnormal data

[0153] The identification of progressive abnormal data adopts the method of multi-machine association and cross-verification. The algorithm process includes: selecting the unit corresponding to the wind speed sequence to be detected as the target unit; selecting multiple adjacent units in the same wind belt as the associated seeds; dividing the wind measurement sequence of the target unit into sequences according to the change point detection algorithm; dividing the wind measurement sequence of the associated seed unit in the same way as the target unit; calculating the wind speed correlation between each subsequence of the target unit and the subsequence at the same position of the associated seed unit one by one; setting the correlation constraint threshold; if the wind speed correlation of a subsequence of the target unit and the seed unit is lower than the threshold and the wind speed correlations between the seed units in the corresponding sequences are all higher than the threshold, then the wind speed subsequence is determined to be abnormal data.

[0154] The process of abnormal wind speed identification method is as follows: Figure 7 shown.

[0155] 5. Example Analysis

[0156] The actual measurement data of wind speed sensor array of a wind farm in my country is taken as the research object. The wind farm is located in the connecting zone between the North China Plain and the Inner Mongolia Plateau. The terrain is flat. Under the influence of the monsoon area and Mongolian high pressure, the duration of winter and summer winds is often more than 7 months. The direction of winter wind is northeasterly. Figure 8 As shown, in this example, unit #6 is selected as the target unit, its abnormal wind measurement data is identified, and units #7 and #8 located upwind are set as the associated seed unit group. The wind measurement data of 10 days from January 1, 2017 to January 10, 2017 are selected as the research object, and the data interval is 10 minutes.

[0157] Calculate T of the wind measurement sequence of Unit #6 h kStatistics, calculation results such as Fig. 9 As shown in the figure, the change point identification is performed according to the binary morphological segmentation algorithm, and the position of the change point is at the peak of the statistic. It can be seen that from January 1, 2017 to January 10, 2017, the number of change points detected is 10, located at the 165th, 254th, 321st, 594th, 730th, 829th, 938th, 1126th, 1231st and 1366th time points respectively.

[0158] The global wind speed sequence is divided into 11 local wind speed subsequences, such as Fig.10 As shown, the subsequence [0,144] is sequence 1, and the subsequences are numbered from left to right as sequence 1 to sequence 11.

[0159] The mean and variance of wind speed of each subsequence are counted. The statistical results are as follows: Fig.11 As shown. It can be seen that the mean and variance of the wind speed of units #7 and #8 in the 11 subsequences are relatively close, and the distribution characteristics are relatively similar; while the average wind speed and variance of unit #6 and units #7 and #8 in subsequence 1 and subsequence 10 are quite different, and are significantly lower than units #7 and #8. In other subsequences, the wind speed statistics of unit #6 and units #7 and #8 are relatively close. In addition, for unit #6, the wind speed means and variances of adjacent subsequences are significantly different, showing different distributions, which proves that the wind speed sequence cutting method established by the present invention can effectively identify the multivariate distribution characteristics of the wind speed sequence, and the division results can reflect the time-varying characteristics of the wind speed sequence.

[0160] Furthermore, the data change speed constraint condition is formulated for each wind speed subsequence using formula (11), and the boundary constraint is calculated according to formula (12). The speed and boundary constraint conditions of each wind speed subsequence are shown in Table 2. The wind speed sequence is subjected to sliding detection with multi-interval constraint conditions. A total of 15 mutation-type abnormal wind measurement data appeared in the interval from January 1, 2017 to January 10, 2017 for Unit #6, indicating that the wind speed sensor communication system of the unit is in good operation and has a relatively stable anti-electromagnetic interference capability. Considering the integrity of the input data for the subsequent wind speed correlation description application, the interpolation algorithm is used in this example to reconstruct the above abnormal data.

[0161] Table 2 Speed ​​and boundary constraints

[0162]

[0163] In order to detect the progressive abnormal data of the target unit in each wind speed subsequence, the t-Copula function is used to calculate the correlation of the wind speeds of the three units in each sequence. The calculation results are as follows: Fig.12As shown, it can be seen that in sequence 1 and sequence 10, the correlation between the measured wind speeds of units #7 and #8 is characterized as highly correlated, with correlation coefficients of 0.91 and 0.86 respectively, while the wind speed correlation between unit #6 and units #7 and #8 is characterized as uncorrelated, with correlation coefficients all below 0.05. The average wind speed of sequence 1 and sequence 10 is statistically found to be significantly lower than that of units #7 and #8, with a difference of nearly 5m / s in average wind speed; in other wind speed subsequences, the three units maintain a high wind speed correlation, with correlation coefficients above 0.6 and up to 0.92. Therefore, it is judged that the wind speed subsequence 1 and subsequence 10 of unit #6 are progressive abnormal wind speeds. From the results of wind speed subsequence division and abnormal wind speed identification, it can be seen that the reason for the change in the distribution characteristics of wind measurement data is not only due to sensor abnormalities, but also due to the time-varying characteristics of the wind speed sequence. In the normal time series interval [165,1231], the wind speed sequence is also divided into multiple different subsequences, and the multivariate statistics between adjacent sequences are quite different. In addition, the wind speed correlation between units also has time-varying characteristics, such as Fig.12 As shown, P(#6,#7) and P(#6,#8) will fluctuate within a certain range in different subsequences, but in each subsequence, a seed unit highly correlated with the wind speed of unit #6 can always be found, and the wind speed correlation coefficient is 0.75 or above. The global wind speed correlation between units #6, #7 and #8 in the entire time series interval is calculated, and the two are 0.617 and 0.639, respectively. It can be seen that the wind speed sequence time series segmentation method used in the present invention has engineering significance for wind speed correlation modeling and analysis, and local wind speed correlation is used as the screening criterion for associated seed units, so that the optimal associated seed unit can always be found in each time series interval, weakening the influence of the time-varying characteristics of wind speed, improving the accuracy of wind speed correlation analysis, and laying a foundation for subsequent abnormal wind speed data identification.

[0164] By calculating the statistical parameters of abnormal data, it is found that in the period from January 1, 2017 to January 10, 2017, the cumulative duration of abnormal wind speed data collected by sensor #6 reached 1650 minutes, the abnormal proportion reached 21.87%, and the root mean square error of abnormal data reached 7.84, the range of deviation from normal data was large, and the average wind speed deviation value was close to 5m / s (relative to the reconstructed wind speed). Among the abnormal data, gradual abnormal data was the main type, mixed with a small amount of sudden abnormal data, indicating that the wind speed sensor data had good anti-electromagnetic interference ability and the wind speed capture module was unstable. On the one hand, it was due to the attenuation of sensor operation performance. It could operate normally in a single and good climate environment, but when the external environmental influencing factors suddenly changed or deteriorated, the sensor operation state was not ideal, and the wind measurement data deviated from the normal wind speed in a large range. Its operation state was largely limited by the influence of external climate and environmental factors; on the other hand, it was due to the attenuation of the performance of the key components of the sensor itself and the unstable operation state, which led to the fluctuation of wind measurement data between normal values ​​and abnormal values. When the performance continued to decline, the proportion of abnormal measured wind speed would increase further.

Claims

1. A method for identifying abnormal wind speed perception data of a wind turbine generator system based on speed association constraints, characterized in that: The following steps are involved: 1) Obtain wind speed sequence data of the target unit; 2) The time series change points of the wind speed sequence are detected by using a binary morphological segmentation algorithm, and the time series change points are used as segmentation points to divide the wind speed sequence into multiple subsequences in time series; 3) Identify abnormal data based on speed and boundary constraints on the divided subsequence data; The abnormal data types include sudden abnormal wind speed data and gradual abnormal wind speed data; In the step 3), the sudden abnormal wind speed data is identified by combining the multi-interval speed constraints and the wind speed upper and lower limit boundary constraints, and the data that does not satisfy the wind data speed change constraint and the wind speed upper and lower limit boundary constraints are determined as sudden abnormal wind speed data. The wind speed upper and lower limit boundary constraints are specifically: in qmin ≤in i ≤in qmax Among them, v qmin and v qmax represent the upper and lower quartiles of wind speed, respectively; The multi-interval speed constraint S is a set of speed constraint intervals. For the rth time window, the speed constraint boundary of the wind speed sequence is determined as follows: Among them, ε u , ε d are sensitivity coefficients, v min and v max Indicates the minimum and maximum wind speeds of wind farm statistics, are the maximum and minimum values ​​of the ascent speed constraint, are the maximum and minimum values ​​of the descent speed constraint respectively.

2. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 1 is characterized in that: The generation situations of the mutation type abnormal wind speed data include: (1) Since the wind speed sensor's ability to resist external electromagnetic disturbances is weakened, when the external disturbance is strong, the measured data will undergo a large-scale mutation, the single abnormality will last for a short time, and the distribution will be discrete; (2) Due to the damage of important parts of the wind speed sensor, the abnormal state undergoes a non-trend mutation, the measured wind speed undergoes a mutation, and the distribution is relatively continuous; The generation of the progressive abnormal wind speed data includes: (1) The degree of damage to the wind speed sensor components is a gradual process, and the measured data deviates from the normal value. The offset gradually increases with the evolution of the abnormal state of the equipment, that is, weak fault data; (2) Consecutive identical values ​​due to communication errors.

3. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 1 is characterized in that: In the step 2), determining the timing change point specifically includes the following steps: 21) Calculate the local statistics of the wind speed sequence, and determine the final change point based on the three constraints of satisfying the local statistic maximum, local peak point and statistic bound; 22) dividing the entire segment of the wind speed sequence data into two segments according to the final change point, and repeating step 21) for each segment to perform segmentation again; 23) Repeat steps 21) and 22) until no new change points appear.

4. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 3 is characterized in that: The step 21) specifically comprises the following steps: 211) For the local data of wind speed sequence V = {V S ,V S+1 ,…,V T }, calculate the corresponding local statistics T h k , k=h+1,…,Th, h represents the search window width, and the local statistics T is obtained h k The point k0 corresponding to the maximum * , and mark this point as a suspected change point; 212) If the suspected change point meets the determination condition of the local peak point, the suspected change point is determined to be a local peak point. The determination condition of the local peak point is specifically: Taking the suspected change point as the center, the points within the window width of 2h are differentiated. If the sum of the differences within the range of h on the left of the suspected change point that are greater than 0 and the differences within the range of h on the right that are less than 0 accounts for the proportion of the window width 2h (k0 * ) is greater than the threshold γ, then the suspected change point is judged to be a local peak point; 213) Define the boundary that distinguishes change points from non-change points a , and according to the boundary a The final changing points are obtained after removing the invariant points from the local peak points.

5. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 4 is characterized in that: In step 211), the local statistic T h k The calculation formula is: Among them, C represents the search coefficient, h represents the search window width, which is related to the sample size T, k represents the point to be detected, and v i 、v j are the i-th and j-th points in the local data respectively.

6. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 5 is characterized in that: In the step 212), the specific gravity P (k0 * ) is: Among them, df(i,j) represents the statistics of two points i and j and The difference of df(i,j) = 1, 0, and -1 respectively indicate that the statistic is increasing, stable, and decreasing. df(i,j)＝1 Indicates that the point is the count of the rising point of the statistic, I df(i,j)＝-1 Indicates that the point is the count of the statistical drop point, and the threshold γ is 0.7 or 0.

8.

7. A method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 6, characterized in that: In step 213), the boundary ξ a Defined as: Among them, a is a parameter used to distinguish non-changing points.

8. The method for identifying abnormal wind speed sensing data of a wind turbine generator system based on speed association constraints according to claim 1, characterized in that: In the step 3), the progressive abnormal wind speed data is identified based on multi-machine association matching, specifically: A wind speed mutual diagnosis community is formed by multiple adjacent units in the same wind belt, and a wind speed sequence correlation description model based on the Copula function is constructed. Multiple units adjacent to the target unit are selected as associated seed units, and the correlation between the target unit and the seed unit group in different wind speed subsequences is calculated one by one. If the correlation between the target unit and the seed unit group in a wind speed subsequence is lower than 0.5 and the correlation between the seed units is higher than 0.75, the wind speed subsequence corresponding to the target unit is judged to be progressive abnormal data. The calculation formula of the correlation is: Among them, P(V1, V2) is the correlation coefficient of the wind speed subsequences V1 and V2 of the target unit and the seed unit, and C(u, v) is the joint probability distribution function.

Images