DIAGNOSTIC AND MONITORING TECHNOLOGIES IN SYSTEM LEVEL FOR WIND TURBINES
Diagnostic and Monitoring Technology Based on Power Curve
The power curve which can describe the relationship between wind speed and electric power output, is a method widely used to evaluate the performance of wind turbines [80-83].
Adaptive neuro-fuzzy interference system (ANFIS) model has been established in [80], for wind turbine power curve detection after considering the parameters of ambient temperature and wind direction. The ANFIS model is specifically constructed as follows: (1) fuzzy interference system (FIS) structure: There are two types of common fuzzy reasoning: Mamdani and Sugeno. Because Sugeno has a fuzzy set involving only the premise part and its computational efficiency is higher, Sugeno type is used as fuzzy inference. (2) Membership functions (MF): The generalized normal distribution MF is used in the input space, and the linear membership function is used as the output space. The advantage of the generalized normal distribution MF is that it can provide a wide range of flexibility in function shape according to the parameters of the function, and can guarantee the smoothness of the transition in the input space. (3) Number of MFs/rules: The number of rules is usually determined by experts familiar with the system to be modelled. In this article, the MF was set to 3 for each input. (4) Training method: Use a mixed learning rule estimated by gradient decency and least squares.
By comparing the root mean squared (RMS), mean absolute error (MAE), mean absolute percentage error (MAPE) and standard deviation (SD) indicators
FIGURE 7.27 Model types used for power output modeling.
of ANFIS, cluster center fuzzy logic (CCFL), neural network (NN) and k-nearest neighbor (K-NN) in the two models, as shown in Figure 7.27, ANFIS not only shows the best performance in this indicator, but also can better detect abnormal power output faults.
In reference [81], an algorithm that can automatically calculate the limit of power curve is proposed. The algorithm automatically generates warning messages when the measured data of wind turbines deviate from the limit of the power curve or hover between the warning area and the warning area. The overall process of setting the limit value of the power curve is shown in Figure 7.28. The first stage is sorted according to the speed-power data measured by speed bin. In the second stage, the average and standard deviation of power are calculated according to each speed bin. The speed bin’s width (unit: m/s) is determined by Equation (7.32).
If the bin’s width is 1 m/s, it becomes “1/2 m/s, l/3m/s ...” with each iteration of the algorithm. The third stage uses interpolation to estimate the power curves. According to the average power obtained in stage 2 as the input of this stage. The fourth stage is to eliminate the abnormal data by moving the estimated power curve left and right or up and down to get the limit of the optimal power curve. In the fifth stage, the data within the limit of the power curve will be taken as input after the abnormal data are excluded. The last stage is to determine whether to terminate the entire algorithm cycle. Calculate the average power standard deviation of each bin obtained in the second phase and compare it with the average power standard deviation of each bin in the entire algorithm cycle before.
FIGURE 7.28 Automatic power curve limit calculation algorithm.
After the limit value of the power curve is obtained, the fault data queue method is used to generate the alert message. The fault data queue consists of three fault data sets. Because the queue has a FIFO data structure, when the measured data enters the warning or alert area, the data set is stored and the earliest failure data set is deleted. If the failure data set queues are all deleted, a warning or alert message for the current activity is released. If the data set in the queue is saturated, a warning or alert message is issued based on the defined conditions.
In [82], an improved Cholesky decomposition Gaussian process (GP) is used to construct a multi-variable power curve model, and longitudinal and transverse data are compared to check the operation of specific components, so as to timely detect the fan operation anomalies and degradation. According to the influence degree of different factors on the output power, wind speed and direction, pitch angle, yaw error, rotor speed and tip speed ratio are selected as the input of the model, and the output of the model is power output prediction. GP modelling is essentially a Bayesian method, which has good performance for random data. A Gaussian Process is completely specified by its mean value m(x) and covariance function k(x, x') and can be written as:
In reference [83], the squared exponential is used as the covariance function. Its matrix form is:
where crj is the signal variance and o;, is the noise variance. D = diag(d{,d2,--,dL) is the length scale parameter for each of the model inputs which links the time variation of the input parameter to that of the output. In this paper, a Cholesky decomposition is used to compute K4 since it is fast and numerically stable. With Cholesky decomposition, the inversion of matrix К can be computed as:
The predicted value of Gaussian Process becomes:
where f is the input of system and y* is unknown predicted output, c is the mean predicted of value of y*.
When a wind turbine encounters a failure, the relationship between power and input factors will deviate from the model, which will cause the model’s predicted residuals to grow, so the sequential probabilistic ratio test (SPRT) is used to detect any abnormal changes in the GP power curve residuals. SPRT consists of two possible testing hypotheses. Hypotheses H0: the wind turbine is fault free and the model residuals have a normal distribution with mean value fi0 and variance <7,7; Hypotheses H{. the wind turbine exhibits abnormal operation with the mean value and variance of the model residuals respectively changing to /л[ and fff respectively. For the GP power curve model residual sequence ex,e2,---,en, the joint probability densities for H0 and W, are respectively as follows:
The SPRT ratio (or the likelihood ratio) is:
The false alarm probability and missed alarm probability respectively are set as a and /3 which give a lower limit A and an upper limit В respectively as:
FIGURE 7.29 Online monitoring procedure of the proposed model.
If R,„ < A, hypothesis #() should be accepted and the wind turbine is regarded as operating normally. Conversely, if Rm > B, hypothesis H0 should be rejected #, accepted instead and the wind turbine operation is regarded as abnormal, triggering an alarm.
In [83], the turbines with weak power generation performance were evaluated by evaluating the wind curve. The least squares method fits the power curve model to the supervisory control and data acquisition (SCADA) data set to construct the curve and shape of the wind curve in continuous time intervals. In order to identify power curves with abnormal curvature and shape, in the next section, a multivariate method for monitoring the contour of the power curve and a residual method for monitoring the power curve fitting error will be introduced. The SCADA data set is divided into sub-data sets according to a fixed time interval, thereby generating a power curve profile and a power curve fitting error. Use multivariate methods and residual methods to develop control charts for monitoring. Continuously collect SCADA data, generate new power curve profiles, and apply developed control charts. The specific implementation process is shown in Figure 7.29.
Since this solution is based on the analysis of sub-data sets, the monitoring effectiveness and accuracy are high. But compared with the point-based state analysis method, this method is not timely, the reason is that it takes some time to construct a new power curve profile. Besides, the relationship between wind energy and multiple parameters was not considered in reference [83].
