Wind turbine generator tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning

Abstract

The invention discloses a wind turbine generator tower vibration early warning method based on a particle swarm optimization algorithm and Mamdani fuzzy reasoning. The wind turbine generator tower vibration early warning method comprises the steps that n variables highly related to wind turbine generator tower vibration are screened out from N variables related to the wind turbine generator tower vibration; calculating the mahalanobis distance of p speed variables in a normal state to form a vector D1; the mahalanobis distance of n-p non-speed variables in the normal state is calculated to form a vector D2, and the vibration state threshold value of the wind turbine generator tower is divided according to the mahalanobis distance; a Mamdani fuzzy reasoning model is established according to the wind turbine generator tower vibration state threshold value, early warning is performed on wind turbine generator tower vibration, and a wind turbine generator vibration early warning model is evaluated by using the precision test index; and optimizing parameters of the Mamdani fuzzy reasoning model to obtain an optimal wind turbine generator tower vibration early warning model. According to the method, the vibration of the wind turbine generator tower can be early warned in advance, the model after optimization improves the abnormal vibration detection rate of the wind turbine generator tower, and the false alarm rate is reduced.

Description

Technical Field

[0001] The invention relates to the technical field of wind turbine fault warning, and in particular to a wind turbine tower vibration warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning. Background Art

[0002] As the main supporting component of the wind turbine, the tower fixes the cabin at a height of tens or even hundreds of meters, so that the wind turbine can better obtain wind energy. During the service life of the wind turbine, the tower will be subjected to complex dynamic loads of different wind conditions, the cabin and its own weight, which will cause abnormal vibration and excessive tilt of the wind turbine tower, which will cause safety incidents of the wind turbine. Therefore, the research on wind turbine tower vibration status monitoring technology can timely detect abnormal vibration problems of wind turbine tower, which is of great significance to improving the economy and safety of wind turbines.

[0003] At present, although there has been some progress in the research on wind turbine tower vibration early warning, many scholars currently choose wind turbine tower vibration early warning models, and the models show a low fault alarm rate and a short advance alarm time. In addition, existing studies often ignore the situation of wind turbine fault false alarms, and do not explore how to reduce the wind turbine fault false alarm rate, thus limiting the early warning accuracy of the model. Summary of the invention

[0004] To solve the above technical problems, the present invention provides a wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning. The method establishes a normal vibration state model of the wind turbine tower based on Mahalanobis distance. The model can determine the wind turbine tower vibration early warning threshold, and combines it with the Mamdani fuzzy reasoning model to propose a wind turbine tower vibration early warning model. The parameters of the Mamdani fuzzy reasoning model are optimized based on three types of early warning indicators, and the output result is obtained through fuzzy optimization reasoning. The method can provide early warning for wind turbine tower vibration, and the optimized model improves the abnormal detection rate of wind turbine tower vibration and reduces the false alarm rate.

[0005] The technical solution adopted by the present invention is:

[0006] The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning includes the following steps:

[0007] Step 1: Select n variables with high correlation with wind turbine tower vibration from N variables related to wind turbine tower vibration, including: p speed variables and np non-speed variables;

[0008] Step 2: Calculate the Mahalanobis distances of p velocity variables under normal conditions to form vector D1; calculate the Mahalanobis distances of np non-velocity variables under normal conditions to form vector D2, and divide the vibration state thresholds of the wind turbine tower according to the Mahalanobis distances; Step 3: Establish a Mamdani fuzzy reasoning model based on the vibration state thresholds of the wind turbine tower to warn of the wind turbine tower vibration, and use the accuracy test index to evaluate the wind turbine vibration warning model;

[0009] Step 4: Optimize the parameters of the Mamdani fuzzy reasoning model to obtain the optimal wind turbine tower vibration warning model.

[0010] The step 1 comprises the following steps:

[0011] S1.1: Collect the SCADA operation data variables of a wind farm every 10 minutes, including wind speed, rotor speed, power, transformer temperature, wind direction, stator temperature, cabin temperature, blade position, rotor bearing temperature, gearbox oil temperature and other variables; as shown in Table 9, it is the SCADA operation data variables of the wind farm every 10 minutes from 0:00 on January 1, 2021 to 23:50 on June 30, 2021.

[0012] Table 9 SCADA operation data variables

[0013]

[0014] S1.2: The above variables are input into the comprehensive correlation index to evaluate their correlation with the tower acceleration, thereby obtaining the comprehensive correlation value of each variable as follows:

[0015] By calculating the comprehensive correlation index of N variables and the X-direction tower acceleration and arranging them in descending order, the most important variables are determined to construct a wind turbine tower early warning model. The corresponding formula is as follows:

[0016]

[0017] Where: η IC is the comprehensive correlation value; r is the Person correlation coefficient; ρ is the Spearman rank correlation coefficient, τ is the Kendall rank correlation coefficient, and I is the maximum information coefficient;

[0018] S1.3: Then, variables with comprehensive correlation values ​​greater than 0.5 are selected and divided into speed variables and non-speed variables. Among them, 5 speed variables are selected: X-direction wind speed, Y-direction wind speed, axial tower acceleration, rotor speed, and generator speed; 5 non-speed variables are selected: power, motor current, front end bearing temperature, rear end bearing temperature, and stator temperature.

[0019] The step 2 comprises the following steps:

[0020] S2.1: By applying the speed variables and non-speed variables selected by the comprehensive correlation index, the normal state Mahalanobis distance is calculated to obtain the speed Mahalanobis distance and non-speed Mahalanobis distance, as follows:

[0021] The normal vibration state of the wind turbine tower is calculated by calculating the Mahalanobis distance of the p velocity variables of the wind turbine tower to represent the normal vibration state of the wind turbine tower. The velocity variables include v, n R , each variable has m historical observations. For the i-th observation, i = 1, 2, ..., m, the standardized value of the mean deviation corresponding to the variable v is j, j = 1, 2, ..., p; where σ is the standard deviation matrix of p speed class variables. Similarly, by calculating the mean standardization of each variable of the p speed class, the vector f shown in the following formula can be obtained i :

[0022]

[0023] In the above formula, the vector f i It consists of p columns of speed variables with normalized mean values; σ j is the standard deviation of the jth column, v i is the wind speed at the i-th moment, is the average wind speed; n Ri is the rotor speed at the i-th moment, is the average rotor speed; a Ti is the tower acceleration at the i-th moment, is the average tower acceleration; n Gi is the generator speed at the i-th moment, is the average generator speed.

[0024] On this basis, the inverse matrix R of the correlation coefficient matrix of speed variables is introduced -1 Integrate p columns of data into a column of dimensionless data, namely the Mahalanobis distance D 1,i , the formula is as follows:

[0025]

[0026] Among them, f i T is the vector f i The transposed matrix of the velocity calculated at m moments is the Mahalanobis distance D 1,1 ,D 1,2 ,…,D 1,m Constitute the vector D1; when the velocity Mahalanobis distance D 1,i The higher the deviation from the mean, the higher the abnormality of the wind turbine tower vibration. Similarly, replace the speed variables in the formula with non-speed variables to calculate the vector f;

[0027] Then, substitute the above formula to introduce the inverse matrix R of the correlation coefficient matrix of non-speed variables -1 , that is, it can calculate the non-velocity Mahalanobis distance D 2,i , calculate the non-velocity Mahalanobis distance D 2,1 ,D 2,2 ,…,D 2,m Constitute vector D2.

[0028] S2.2: The above-mentioned speed Mahalanobis distance and non-speed Mahalanobis distance are divided into different vibration state thresholds of wind turbine tower using the “3σ” criterion of normal distribution, as follows:

[0029] The "3σ" criterion of normal distribution is used to divide the right side of the frequency histogram of the speed Mahalanobis distance D1 into three different vibration state thresholds of the wind turbine tower. Among them, the normal vibration state threshold is μ1+σ1, the medium vibration threshold is μ1+2σ1, and the vibration abnormality threshold is μ1+3σ1, μ1 is the mean of the speed Mahalanobis distance D1, and σ1 is the standard deviation of the speed Mahalanobis distance D1.

[0030] Similarly, three different vibration state thresholds are divided according to the frequency histogram of the non-speed Mahalanobis distance D2: the normal vibration state threshold is μ2+σ2, the medium vibration threshold is μ2+2σ2, and the vibration abnormality threshold is μ2+3σ2, where μ2 is the mean of the speed Mahalanobis distance D2, and σ2 is the standard deviation of the speed Mahalanobis distance D2.

[0031] The step 3 comprises the following steps:

[0032] S3.1: Use the different vibration state thresholds of the wind turbine tower to establish the Mamdani fuzzy reasoning model, take the speed Mahalanobis distance and non-speed Mahalanobis distance as input, and the different vibration states of the wind turbine as output, construct the corresponding membership function, establish a rule base, and use the centroid method to convert the output into a numerical value, and output a line graph to provide a wind turbine tower vibration warning; the details are as follows:

[0033] The velocity Mahalanobis distance D1 is used as input, and the vibration state of the wind turbine tower is divided into three fuzzy sets according to the calculated vibration threshold: normal vibration state β L1 , medium vibration state β M1 , abnormal vibration state β H1 , and construct the membership functions respectively; similarly, the Mahalanobis distance D2 calculated by non-speed variables is also divided into three fuzzy sets to construct the corresponding membership functions;

[0034] Through the fuzzy sets of two input data, a rule base for the vibration state of the wind turbine tower is established to judge the vibration state of the wind turbine tower. The rule base contains 9 inference rules.

[0035] Fuzzy output U = {normal vibration, medium vibration, abnormal vibration}, output U is obtained by the AND operation of each membership function, the formula is as follows:

[0036]

[0037] In the formula, ∪ represents the union; g(u) is the output of fuzzy reasoning, It is the minimum operation of the input fuzzy set.

[0038] Apply the centroid method to take the centroid of the membership function curve and the area enclosed by the horizontal coordinate, and divide the output membership function into 5 areas, with areas A1, A2, ..., A5 respectively. The horizontal coordinates corresponding to the centroid of each area are expressed as c1, c2, ..., c5. According to the following formula, the defuzzified output value u can be obtained c .

[0039]

[0040] S3.2: Use three accuracy test indicators, namely, vibration anomaly detection rate, false alarm rate, and average timely warning time, to evaluate the performance of the wind turbine tower vibration warning model; specifically, include:

[0041] The vibration anomaly detection rate is the number of times the wind turbine tower vibration threshold T is exceeded within the abnormal range. h The ratio of the number of events to the total number of abnormal wind turbine tower vibration events is calculated as follows:

[0042]

[0043] In the formula, is the vibration abnormality detection rate, H d The vibration score is greater than T in the abnormal vibration state h The number of events.

[0044] The false alarm rate is defined as: the wind turbine tower exceeds the wind turbine tower vibration warning threshold T in normal operation. h The false alarm rate is calculated by the ratio of the number of events in the normal operating state to the total number of events in the normal operating state, and the SCADA data of group T2 containing the normal operating wind turbine tower is used, as shown in the following formula.

[0045]

[0046] In the formula, is the false alarm rate, H q The vibration fraction is greater than T in normal stateh The number of events.

[0047] The average timely warning time is defined as: the mean of the difference between the wind turbine tower vibration warning time and the wind turbine operation log alarm time, as shown in the following formula.

[0048]

[0049] In the formula, is the average timely warning time, T e is the vibration alarm time of the running log, T n It is the tower vibration warning time.

[0050] In step 4, the parameters of the Mamdani fuzzy inference model are optimized to obtain the optimal wind turbine tower vibration early warning model; the steps include:

[0051] S4.1: Particle swarm initialization Set the particle swarm number N to 5, the swarm particle velocity boundary to [-1,1], and use random numbers to make the generated particle initial position and particle velocity within the range of [0,1]. The boundary range corresponding to the particles is shown in the following formula.

[0052]

[0053] Where μ1 is the mean of the Mahalanobis distance D1, μ2 is the mean of the Mahalanobis distance D2, σ1 is the standard deviation of the Mahalanobis distance D1, and σ2 is the standard deviation of the Mahalanobis distance D2.

[0054] S4.2: The function value of each particle is calculated through the objective function, and the maximum wind turbine tower vibration warning index is used as the optimization target. The parameters b1, c1, e1, b2, c2, and e2 of the membership function are encoded into the particle group for global adjustment. After the parameter optimization is determined, the optimization model is as follows:

[0055]

[0056] In the formula, is the vibration abnormality detection rate, is the false alarm rate, To provide timely warning time, is the expected warning time, set to 5.

[0057] S43: Substitute the initial particle position into the calculated value of the objective function, and then update the particle velocity and particle position. In the process of finding the optimal value of the particle, the particle velocity is updated as shown in the following formula:

[0058]

[0059] In the above formula, v i(t+1) is the velocity of the ith particle at time t+1, w(t) represents the inertia factor at time t, and v i (t) is the velocity of particle i; and are the particle self-learning coefficient and particle swarm learning coefficient, which are set to 1.5 and 2 respectively; r1 and r2 are random numbers ranging from 0 to 1; pbest i is the individual optimal value, x i is the current position of the particle, gbest is the global optimal value;

[0060] The particle position is given by:

[0061] x i (t+1)=x i (t)+v i (t+1)

[0062] In the formula, x i (t+1) is the position of the particle at time t+1, x i (t) is the position of the particle at time t. Compare it with the historical individual optimal value and determine whether to update the individual optimal position; then find the group optimal value based on the individual optimal value, compare it with the current group optimal value and determine whether to update the group optimal position.

[0063] S44: Repeat the steps in S42 and S43. When the maximum number of iterations reaches 100, the iteration ends and the global optimal result is output.

[0064] The present invention provides a wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning, and the technical effects are as follows:

[0065] 1) Step 1 of the present invention can select variables with strong correlation with the vibration of the wind turbine tower from a large number of SCADA data variables by calculating the comprehensive correlation value, so that the subsequent calculation of the Mahalanobis distance can more accurately reflect the vibration of the wind turbine tower.

[0066] 2) Step 2 of the present invention can determine the vibration state of the wind turbine tower by calculating the Mahalanobis distance, that is, calculating the degree of difference between multiple groups of variables and the average values ​​of all observed values ​​of their variables, and divide the wind turbine tower into different thresholds according to different numerical ranges of the Mahalanobis distance to characterize the vibration state of the wind turbine tower.

[0067] 3) Step 3 of the present invention uses fuzzy reasoning to fuzzify the threshold boundary, judge the degree to which the Mahalanobis distance belongs to different vibration states, and then obtain the output value through Mamdani fuzzy reasoning, so as to judge whether the wind turbine tower has abnormal vibration, and to a certain extent avoid the misjudgment problem caused by the wind turbine tower vibration exceeding the set threshold for a short time.

[0068] 4) Step 4 of the present invention optimizes the fuzzy reasoning parameter values ​​to make the fuzzy reasoning model results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0070] Figure 1 The present invention takes into account the flow chart of the wind turbine tower vibration early warning model.

[0071] Figure 2 Schematic diagram of Mamdani fuzzy reasoning model.

[0072] Figure 3(a) is the frequency histogram of the velocity-type Mahalanobis distance.

[0073] Figure 3(b) is the frequency histogram of non-velocity Mahalanobis distance.

[0074] Figure 4 This is the output line graph of the abnormal vibration samples of the Penmanshiel wind turbine.

[0075] Figure 5 This is the output line detail diagram of the abnormal vibration sample of the Penmanshiel wind turbine.

[0076] Figure 6 This is the output line graph of the normal state sample of the Penmanshiel wind turbine measured. DETAILED DESCRIPTION

[0077] A wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning, firstly, based on the data of wind turbine supervisory control and data acquisition system, the velocity Mahalanobis distance and non-velocity Mahalanobis distance of normal vibration state are calculated by using variables with high correlation with wind turbine tower vibration, and three different vibration state thresholds are determined respectively; secondly, the Mamdani fuzzy reasoning model is established to divide the fuzzy sets according to different vibration state thresholds, and the membership functions are established respectively; finally, the particle swarm optimization algorithm is used to optimize the parameters of the membership function, and the fuzzy optimization reasoning model is obtained, and the output numerical line graph is used to judge whether the wind turbine tower vibration alarm is triggered. This model provides a scientific and accurate basis for wind turbine tower vibration early warning.

[0078] like Figure 1 As shown, a wind turbine tower vibration early warning method comprises the following steps:

[0079] S1: Use comprehensive correlation index to select n variables with high correlation with wind turbine tower vibration from variables that may be related to wind turbine tower vibration.

[0080] S11: By calculating the comprehensive correlation index of N variables and the X-direction tower acceleration and arranging them in descending order, the most important variables are determined to construct a wind turbine tower warning model. The corresponding formula is as follows:

[0081]

[0082] Where: r (|r|≤1) is the Person correlation coefficient; ρ (|ρ|≤1) is the Spearman rank correlation coefficient, τ (|τ|≤1) is the Kendall rank correlation coefficient, and I (0≤I≤1) is the maximum information coefficient.

[0083] S12: Arrange the comprehensive correlation indexes in descending order, and select n (n≤N) variables whose comprehensive correlation indexes are greater than 0.5, among which there are p speed variables and np non-speed variables;

[0084] S2: Calculate the speed-type and non-speed-type Mahalanobis distances in the normal state, and divide the different vibration state thresholds of the wind turbine according to the Mahalanobis distance.

[0085] S21: Apply p velocity variables of the wind turbine tower to calculate the Mahalanobis distance of the normal vibration state to represent the normal vibration state of the wind turbine tower. According to numerical experiments, velocity variables usually include v, n R , each variable has m historical observations. For the i-th (i=1, 2, …, m) observation, the standardized value of the mean deviation corresponding to the variable v is (j=1, 2, …, p), where σ is the standard deviation matrix of the p speed class variables. Similarly, the mean standardization of each variable of the p speed class is calculated, and the vector f shown in the following formula can be obtained. i :

[0086]

[0087] Vector f i It consists of p columns of speed variables with normalized mean values; σ j is the standard deviation of the jth column, v i is the wind speed at the i-th moment, is the average wind speed; n Ri is the rotor speed at the i-th moment, is the average rotor speed; a Ti is the tower acceleration at the i-th moment, is the average tower acceleration; n Gi is the generator speed at the i-th moment, is the average generator speed.

[0088] S22: On this basis, the inverse matrix R of the correlation coefficient matrix of speed variables is introduced -1Integrate p columns of data into a column of dimensionless data, namely the Mahalanobis distance D 1,i , the formula is as follows:

[0089]

[0090] Among them, f i T is the vector f i The transposed matrix of the velocity calculated at m moments is D 1,1 ,D 1,2 ,…,D 1,m Constitute the vector D1. When the velocity Mahalanobis distance D 1,i The higher the deviation from the mean, the higher the abnormality of the wind turbine tower vibration. Similarly, replace the speed variables in the formula with non-speed variables to calculate the vector f; then, substitute the inverse matrix R of the correlation coefficient matrix of non-speed variables into the above formula. -1 , then the non-velocity Mahalanobis distance D can be calculated 2,i , calculate the non-velocity Mahalanobis distance D 2,1 ,D 2,2 ,…,D 2,m Constitute vector D2.

[0091] S23: The "3σ" criterion of normal distribution is used to divide the right side of the frequency histogram of the speed-type Mahalanobis distance D1 into three different vibration state thresholds for the wind turbine tower. Among them, the normal vibration state threshold is μ1+σ1, the medium vibration threshold is μ1+2σ1, and the vibration abnormality threshold is μ1+3σ1. μ1 is the mean of the speed-type Mahalanobis distance D1, and σ1 is the standard deviation of the speed-type Mahalanobis distance D1. Similarly, three different vibration state thresholds are divided according to the frequency histogram of the non-speed-type Mahalanobis distance D2: the normal vibration state threshold is μ2+σ2, the medium vibration threshold is μ2+2σ2, and the vibration abnormality threshold is μ2+3σ2. μ2 is the mean of the speed-type Mahalanobis distance D2, and σ2 is the standard deviation of the speed-type Mahalanobis distance D2.

[0092] S3: The Mamdani fuzzy inference system is combined with Mahalanobis distance to construct a wind turbine tower vibration model, and three accuracy test indicators are used for evaluation.

[0093] S31: Figure 2 The Mamdani type fuzzy inference model takes the velocity Mahalanobis distance D1 as input and divides the wind turbine tower vibration state into three fuzzy sets according to the calculated vibration threshold: normal vibration state (β L1 ), medium vibration state (β M1 )、Abnormal vibration state (β H1), and construct the membership function respectively. The present invention comprises the triangular membership function and the trapezoidal membership function to form the membership function. Similarly, the present invention also divides the Mahalanobis distance D2 calculated by the non-speed variable into three fuzzy sets to construct the corresponding membership function.

[0094] S32: through the fuzzy sets of two input data, a rule base for dividing the vibration state of the wind turbine tower is established as shown in Table 1, and then the vibration state of the wind turbine tower is judged. The rule base contains 9 inference rules.

[0095] Table 1 Rule base

[0096]

[0097]

[0098] S33: Fuzzy output U = {normal vibration, medium vibration, abnormal vibration}. The output U is obtained by the AND operation of each membership function, and the formula is as follows:

[0099]

[0100] In the formula, ∪ represents the union, g(u) is the output of fuzzy reasoning, It is the minimum operation of the input fuzzy set.

[0101] S34: Finally, the centroid method is applied to obtain the centroid of the membership function curve and the area enclosed by the horizontal coordinate, and the output membership function is divided into 5 areas, with areas A1, A2, ..., A5 respectively. The horizontal coordinates corresponding to the centroid of each area are expressed as c1, c2, ..., c5. The defuzzified output value u can be obtained according to the following formula c .

[0102]

[0103] Among them, the vibration warning threshold T h Set to 70.

[0104] S35: Apply three types of wind turbine tower vibration accuracy inspection indicators, namely vibration anomaly detection rate, false alarm rate, and average timely warning time to evaluate model performance. Offline testing is performed using SCADA data of 720 moments of wind turbine tower vibration anomalies in the T1 group and 720 moments of wind turbine tower normal operation in the T2 group.

[0105] The vibration anomaly detection rate is within the abnormal range and exceeds the wind turbine tower vibration threshold T h The ratio of the number of events to the total number of abnormal wind turbine tower vibration events is calculated as follows:

[0106]

[0107] In the formula, is the vibration abnormality detection rate, H d The vibration score is greater than T in the abnormal vibration state h The number of events.

[0108] The false alarm rate is defined as the number of times the wind turbine tower exceeds the wind turbine tower vibration warning threshold T in normal operation. h The present invention uses the SCADA data of group T2 containing wind turbine towers in normal operation to calculate the false alarm rate, as shown in the following formula.

[0109]

[0110] In the formula, is the false alarm rate, H q The vibration fraction is greater than T in normal state h The number of events.

[0111] The average timely warning time is defined as the mean of the difference between the wind turbine tower vibration warning time and the wind turbine operation log alarm time of the model of the present invention, as shown in the following formula.

[0112]

[0113] In the formula, is the average timely warning time, T e is the vibration alarm time of the running log, T n The present invention proposes a model tower vibration warning time.

[0114] S4: Apply the particle swarm optimization algorithm to optimize the parameters of the Mamdani fuzzy inference model in step S3 so that the model effect is optimal.

[0115] S41: Particle swarm initialization sets the particle swarm number N to 5, the swarm particle velocity boundary to [-1,1], and the generated particles' initial positions and particle velocities are set to be in the range of [0,1] by using random numbers. The particle boundary range is shown in the following formula.

[0116]

[0117] Where μ1 is the mean of the Mahalanobis distance D1, μ2 is the mean of the Mahalanobis distance D2, σ1 is the standard deviation of the Mahalanobis distance D1, and σ2 is the standard deviation of the Mahalanobis distance D2.

[0118] S42: Calculate the function value of each particle through the objective function, take the maximum wind turbine tower vibration warning index as the optimization target, encode the parameters b1, c1, e1, b2, c2, e2 of the membership function into the particle group, make global adjustments, and after the parameter optimization is determined, the optimization model is as follows:

[0119]

[0120] In the formula, is the vibration abnormality detection rate, is the false alarm rate, To provide timely warning time, is the expected warning time, set to 5.

[0121] S43: Substitute the initial particle position into the calculated value of the objective function, and then update the particle velocity and particle position. In the process of finding the optimal value of the particle, the particle velocity is updated as shown in the following formula:

[0122]

[0123] In the above formula, v i (t+1) is the velocity of the ith particle at time t+1, w(t) represents the inertia factor at time t, and v i (t) is the velocity of particle i; and are the particle self-learning coefficient and particle swarm learning coefficient, which are set to 1.5 and 2 respectively; r1 and r2 are random numbers ranging from 0 to 1; pbest i is the individual optimal value, x i is the current position of the particle, gbest is the global optimal value;

[0124] The particle position is given by:

[0125] x i (t+1)=x i (t)+v i (t+1)

[0126] In the formula, x i (t+1) is the position of the particle at time t+1, x i (t) is the position of the particle at time t. Compare it with the historical individual optimal value and determine whether to update the individual optimal position; then find the group optimal value based on the individual optimal value, compare it with the current group optimal value and determine whether to update the group optimal position.

[0127] S44: Repeat the steps in S42 and S43. When the maximum number of iterations reaches 100, the iteration ends and the global optimal result is output.

[0128] Example:

[0129] This embodiment collects SCADA data samples measured every 10 minutes from 2018 to 2020 for the Penmanshiel02 wind turbine at the Penmanshiel wind farm, substitutes the data into fuzzy optimization reasoning to obtain the output value and then judge whether abnormal vibration occurs in the tower of the wind turbine.

[0130] First, the SCADA data of normal vibration state of Penmanshiel02 wind turbine from January to June 2018 were selected to calculate the comprehensive correlation index of each variable, and the variables with comprehensive correlation index values ​​greater than 0.5 were selected. The variables selected by the present invention and the comprehensive correlation index values ​​are shown in Table 2. Among them, there are 5 speed variables: X-direction wind speed, Y-direction wind speed, axial tower acceleration, rotor speed, generator speed; 5 non-speed variables: power, motor current, front bearing temperature, rear bearing temperature, stator temperature.

[0131] Table 2 Selected variables and comprehensive related index values

[0132]

[0133] The speed Mahalanobis distance D1 and the non-speed Mahalanobis distance D2 are calculated respectively. The frequency histogram of the speed Mahalanobis distance D1 and the frequency histogram of the non-speed Mahalanobis distance D2 are shown in Figure 3(a) and Figure 3(b). According to the speed Mahalanobis distance D1 of the wind turbine, its mean value μ1=5.32×10 -12 , standard deviation σ1 = 1.00, calculate the three different state thresholds of the wind turbine tower divided by the speed-type Mahalanobis distance D1: the normal vibration state threshold is μ1 + σ1, the medium vibration state threshold is μ1 + 2σ1, the abnormal vibration state threshold is μ1 + 3σ1, and the non-speed-type Mahalanobis distance D2 estimates its mean μ2 = -3.04 × 10 -14 , standard deviation σ2=1.00, and non-velocity Mahalanobis distance D2 divide the wind turbine tower into three different state thresholds: the normal vibration state threshold is μ2+σ2, the medium vibration state threshold is μ2+2σ2, and the abnormal vibration state threshold is μ2+3σ2.

[0134] The speed Mahalanobis distance D1 and the non-speed Mahalanobis distance D2 are divided into three different states of fuzzy sets, and the corresponding membership functions are constructed. The parameter values ​​of the membership functions are shown in Table 3.

[0135] Table 3 Membership function parameter values

[0136]

[0137] The operation log of the Penmanshiel02 wind turbine from July 2018 to June 2020 was consulted, and 10 sets of data records of wind turbine tower vibration shutdown events were selected as shown in Table 4. The selected data includes the SCADA data on the day of the vibration shutdown and the previous four days.

[0138] Table 4 Data records of 10 wind turbine tower vibration shutdown events

[0139]

[0140] In addition, 10 sets of data on normal vibration status of wind turbine towers for five days are selected and shown in Table 5. The empty set symbol in the table indicates that no vibration shutdown event occurred in the wind turbine tower.

[0141] Table 5 Data records of normal operation of 10 wind turbine towers

[0142]

[0143] The parameters are optimized based on the precision test index and the particle swarm optimization algorithm. The initial value of the algorithm's duty cycle is evenly distributed within its search space. Among them, the particle population size N = 5, the particle self-learning coefficient Particle swarm learning coefficient The parameter values ​​after optimization by particle swarm algorithm are shown in Table 6.

[0144] Table 6 Membership function parameter values ​​after particle swarm optimization

[0145]

[0146] The three accuracy test indicators before and after optimization are shown in Table 7. It can be seen from the table that the optimized fuzzy reasoning further improves the abnormal detection rate of wind turbine tower vibration, reduces the false alarm rate, and advances the average timely warning time.

[0147] Table 7 Precision test indicators before and after particle swarm algorithm optimization

[0148]

[0149] The SCADA data of the Penmanshiel02 wind turbine from January to June 2021 were selected and the model was applied for online monitoring. Among them, scenario 1 includes a wind turbine tower vibration shutdown event, and scenario 2 does not have a wind turbine tower vibration shutdown event. The time periods of the two selected scenarios were not used for offline training of the fuzzy optimization reasoning model, so the data is suitable for testing the effectiveness of the model of the present invention for wind turbine tower vibration warning.

[0150] (1) Wind turbine tower vibration shutdown scenario and status monitoring:

[0151] The scenario description including the wind turbine tower vibration shutdown event is shown in Table 8.

[0152] Table 8 Description of wind turbine tower vibration shutdown scenarios

[0153]

[0154] By using the fuzzy optimization reasoning proposed by the present invention, we can get Figure 4 The figure shows the numerical line graph output by the method of the present invention for the Penmanshiel02 wind turbine from 0:00 on February 15, 2021 to 0:00 on March 10, 2021. Figure 4 It can be seen that the model of the present invention had three short-term false alarms from February 15 to March 9, 2021, and exceeded the warning threshold for a long time on March 9.

[0155] Figure 5 Shown Figure 4 The black box contains the time period of March 9, 00:00-23:50, and outputs the numerical line detail graph. Figure 5 It can be seen that the model exceeded the threshold and issued an early warning at 17:30 on March 9, 2021, and exceeded the early warning threshold until 20:50. Compared with the alarm time in the operation log, this shows that the early warning model was about 3.33 hours ahead of time.

[0156] (2) Normal vibration scenario and status monitoring of wind turbine tower:

[0157] The Penmanshiel02 wind turbine at the Penmanshiel wind farm did not experience any wind turbine tower vibration shutdown event from May 20 to June 14, 2021, and was in a normal vibration state. Using the fuzzy optimization reasoning proposed in this description, we can get the following: Figure 6 The following is a line chart showing the output values ​​of the Penmanshiel02 wind turbine status from 0:00 on May 20, 2021 to 0:00 on June 14, 2021. Figure 6 It can be seen that the model had no false alarms from May 26 to June 11, 2021, two false alarms occurred from May 23 to 25, and two false alarms occurred on June 12.

[0158] From the results of the above embodiments, it can be seen that the present invention can be used to provide early warning for wind turbine tower vibration. The optimized model improves the detection rate of wind turbine tower vibration anomalies, reduces the false alarm rate, and provides vibration warnings 2.47 hours in advance on average. The online application results show that the vibration warning can be issued 3.33 hours in advance for scenarios with abnormal wind turbine tower vibration, with 3 false alarms. The false alarm rate of the model is 4 times in normal wind turbine tower vibration scenarios, and no false alarms occurred from May 26 to June 11.

Claims

1. A wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning is characterized by The following steps are involved: Step 1: Select n variables with high correlation with wind turbine tower vibration from N variables related to wind turbine tower vibration, including: p speed variables and np non-speed variables; Step 2: Calculate the Mahalanobis distances of p velocity variables under normal conditions to form a vector D1; calculate the Mahalanobis distances of np non-velocity variables under normal conditions to form a vector D2, and divide the vibration state thresholds of the wind turbine tower according to the Mahalanobis distances; Step 3: Establish a Mamdani fuzzy reasoning model based on the vibration state threshold of the wind turbine tower to warn the wind turbine tower vibration, and use the accuracy test index to evaluate the wind turbine vibration warning model; Step 4: Optimize the parameters of the Mamdani fuzzy reasoning model to obtain the optimal wind turbine tower vibration warning model.

2. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 1 is characterized in that: The step 1 comprises the following steps: S1.1: Collect SCADA operation data variables of a wind farm every 10 minutes, including wind speed, rotor speed, power, transformer temperature, wind direction, stator temperature, nacelle temperature, blade position, rotor bearing temperature, and gearbox oil temperature variables; S1.2: Input the above variables into the comprehensive correlation index to evaluate their correlation with tower acceleration, so as to obtain the comprehensive correlation value of each variable, as follows: By calculating the comprehensive correlation index of N variables and the X-direction tower acceleration and arranging them in descending order, the most important variables are determined to construct a wind turbine tower early warning model. The corresponding formula is as follows: Where: η IC is the comprehensive correlation value; r is the Person correlation coefficient; ρ is the Spearman rank correlation coefficient, τ is the Kendall rank correlation coefficient, and I is the maximum information coefficient; S1.3: Then, variables with comprehensive correlation values ​​greater than 0.5 are selected and divided into speed variables and non-speed variables. Among them, 5 speed variables are selected: X-direction wind speed, Y-direction wind speed, axial tower acceleration, rotor speed, and generator speed; 5 non-speed variables are selected: power, motor current, front end bearing temperature, rear end bearing temperature, and stator temperature.

3. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 1 is characterized in that: The step 2 comprises the following steps: S2.1: By applying the speed variables and non-speed variables selected by the comprehensive correlation index, the normal state Mahalanobis distance is calculated to obtain the speed Mahalanobis distance and the non-speed Mahalanobis distance; S2.2: The above-mentioned speed-related Mahalanobis distance and non-speed-related Mahalanobis distance are divided into different vibration state thresholds of wind turbine tower using the "3σ" criterion of normal distribution.

4. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 3 is characterized by: In S2.1, the normal vibration state Mahalanobis distance is calculated for the p velocity variables of the wind turbine tower to represent the normal vibration state of the wind turbine tower. The velocity variables include v, n R , each variable has m historical observations. For the i-th observation, i = 1, 2, ..., m, the standardized value of the mean deviation corresponding to the variable v is j, j = 1, 2, ..., p; where σ is the standard deviation matrix of p speed class variables. Similarly, by calculating the mean standardization of each variable of the p speed class, the vector f shown in the following formula can be obtained i : In the above formula, the vector f i It consists of p columns of speed variables with normalized mean values; σ j is the standard deviation of the jth column, v i is the wind speed at the i-th moment, is the average wind speed; n Ri is the rotor speed at the i-th moment, is the average rotor speed; a Ti is the tower acceleration at the i-th moment, is the average tower acceleration; n Gi is the generator speed at the i-th moment, is the average generator speed; On this basis, the inverse matrix R of the correlation coefficient matrix of speed variables is introduced -1 Integrate p columns of data into a column of dimensionless data, namely the Mahalanobis distance D 1,i , the formula is as follows: Among them, f i T is the vector f i The transposed matrix of the velocity calculated at m moments is the Mahalanobis distance D 1,1 ,D 1,2 ,…,D 1,m Constitute the vector D1; when the velocity Mahalanobis distance D 1,i The higher the deviation from the mean, the higher the abnormality of the wind turbine tower vibration. Similarly, the speed variables in the formula are replaced with non-speed variables to calculate the vector f. Then, substitute the above formula to introduce the inverse matrix R of the correlation coefficient matrix of non-speed variables -1 , that is, it is possible to calculate the non-velocity Mahalanobis distance D 2,i , calculate the non-velocity Mahalanobis distance D 2,1 ,D 2,2 ,…,D 2,m Constitute vector D2.

5. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 3 is characterized by: In S2.2, the "3σ" criterion of normal distribution is cited to divide the frequency histogram of the speed Mahalanobis distance D1 into three different vibration state thresholds of the wind turbine tower; among them, the normal vibration state threshold is μ1+σ1, the medium vibration threshold is μ1+2σ1, and the vibration abnormality threshold is μ1+3σ1, μ1 is the mean of the speed Mahalanobis distance D1, and σ1 is the standard deviation of the speed Mahalanobis distance D1; Similarly, three different vibration state thresholds are divided according to the frequency histogram of the non-speed Mahalanobis distance D2: the normal vibration state threshold is μ2+σ2, the medium vibration threshold is μ2+2σ2, and the vibration abnormality threshold is μ2+3σ2, where μ2 is the mean of the speed Mahalanobis distance D2, and σ2 is the standard deviation of the speed Mahalanobis distance D2.

6. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 1 is characterized by: The step 3 comprises the following steps: S3.1: Use the different vibration state thresholds of the wind turbine tower to establish the Mamdani fuzzy reasoning model, take the speed Mahalanobis distance and non-speed Mahalanobis distance as input, and the different vibration states of the wind turbine as output, construct the corresponding membership function, establish a rule base, and use the centroid method to convert the output into a numerical value, and use the output line graph to perform wind turbine tower vibration warning; S3.2: Use three accuracy test indicators, namely vibration anomaly detection rate, false alarm rate, and average timely warning time, to evaluate the performance of the wind turbine tower vibration warning model.

7. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 6 is characterized by: In S3.1, the velocity Mahalanobis distance D1 is used as input, and the vibration state of the wind turbine tower is divided into three fuzzy sets according to the calculated vibration threshold: normal vibration state β L1 , medium vibration state β M1 , abnormal vibration state β H1 , and construct the membership functions respectively; similarly, the Mahalanobis distance D2 calculated by non-speed variables is also divided into three fuzzy sets to construct the corresponding membership functions; Through the fuzzy sets of two input data, a rule base for the vibration state of the wind turbine tower is established to judge the vibration state of the wind turbine tower. The rule base contains 9 inference rules. Fuzzy output U = {normal vibration, medium vibration, abnormal vibration}, output U is obtained by the AND operation of each membership function, the formula is as follows: In the formula, ∪ represents the union; g(u) is the output of fuzzy reasoning, It is the operation of taking the minimum of the input fuzzy set; Apply the centroid method to take the centroid of the membership function curve and the area enclosed by the horizontal coordinate, and divide the output membership function into 5 areas, with areas A1, A2, ..., A5 respectively. The horizontal coordinates corresponding to the centroid of each area are expressed as c1, c2, ..., c5. According to the following formula, the defuzzified output value u can be obtained c ; 8. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 6 is characterized by: In S3.2, the vibration anomaly detection rate is the number of wind turbine tower vibrations exceeding the wind turbine tower vibration threshold T within the abnormal range. h The ratio of the number of events to the total number of abnormal wind turbine tower vibration events is calculated as follows: In the formula, is the vibration abnormality detection rate, H d The vibration score is greater than T in the abnormal vibration state h The number of events; The false alarm rate is defined as: the wind turbine tower exceeds the wind turbine tower vibration warning threshold T in normal operation. h The ratio of the number of events to the total number of events in normal operation is used to calculate the false alarm rate using the SCADA data of group T2 containing wind turbine towers in normal operation, as shown in the following formula; In the formula, is the false alarm rate, H q The vibration fraction is greater than T in normal state h The number of events; The average timely warning time is defined as: the mean of the difference between the wind turbine tower vibration warning time and the wind turbine operation log alarm time, as shown in the following formula; In the formula, is the average timely warning time, T e is the vibration alarm time of the running log, T n It is the tower vibration warning time.

9. The wind turbine tower vibration early warning method based on particle swarm optimization algorithm and Mamdani fuzzy reasoning according to claim 1 is characterized by: In step 4, the parameters of the Mamdani fuzzy inference model are optimized to obtain the optimal wind turbine tower vibration early warning model; the steps include: S4.1: Particle swarm initialization Set the particle swarm number N to 5, the swarm particle velocity boundary to [-1,1], and use random numbers to make the generated particle initial position and particle velocity in the range of [0,1]. The boundary range corresponding to the particles is shown in the following formula; Where, μ1 is the mean of the Mahalanobis distance D1, μ2 is the mean of the Mahalanobis distance D2, σ1 is the standard deviation of the Mahalanobis distance D1, and σ2 is the standard deviation of the Mahalanobis distance D2; S4.2: The function value of each particle is calculated through the objective function, and the maximum wind turbine tower vibration warning index is used as the optimization target. The parameters b1, c1, e1, b2, c2, and e2 of the membership function are encoded into the particle group for global adjustment. After the parameter optimization is determined, the optimization model is as follows: In the formula, is the vibration abnormality detection rate, is the false alarm rate, To provide timely warning time, is the expected warning time; S43: Substitute the initial particle position into the calculated value of the objective function, and then update the particle velocity and particle position. In the process of finding the optimal value of the particle, the particle velocity is updated as shown in the following formula: In the above formula, v i (t+1) is the velocity of the ith particle at time t+1, w(t) represents the inertia factor at time t, and v i (t) is the velocity of particle i; and are particle self-learning coefficient and particle swarm learning coefficient respectively; r1 and r2 are random numbers from 0 to 1 respectively; pbest i is the individual optimal value, x i is the current position of the particle, gbest is the global optimal value; The particle position is given by: x i (t+1)=x i (t)+v i (t+1) In the formula, x i (t+1) is the position of the particle at time t+1, x i (t) is the position of the particle at time t; compare it with the historical individual optimal value and determine whether to update the individual optimal position; then find the group optimal value based on the individual optimal value, compare it with the current group optimal value and determine whether to update the group optimal position; S44: Repeat S42 and S43. When the maximum number of iterations is reached, the iteration ends and the global optimal result is output.

Images