Comparison of two state evaluation methods for wind power generation equipment

The existing intelligent warning system achieved intelligent warning through state prediction, state evaluation, and fault localization. This article proposed and compared two state evaluation methods based on adjustable smoothing parameters and sliding window similarity. Actual wind turbine data was used as training data for the multivariate state prediction model, and the predicted data was used as input for the two state judgment methods. It was concluded that the sliding window similarity method has advantages in early warning accuracy, timeliness, and simplicity. Finally, the influence of sliding window parameters on early warning sensitivity was analyzed and discussed.


Introduction
The operation status of wind power generation equipment depends on the overall structure of the wind power generation equipment.By evaluating the operation status of wind power generation equipment, we can timely discover hidden dangers and provide a basis for taking corresponding maintenance and repair measures for the faulty parts of the equipment.This can not only save the maintenance and repair costs of wind power generation equipment but also, importantly, prevent the entire power generation process from stopping due to structural failures, which may even endanger people's lives and cause significant economic losses to enterprises and the country.
There are many methods for evaluating the status of wind power generation equipment, which can be broadly divided into mechanism modeling methods and data-driven intelligent methods.In the process of mechanism modeling, inputs, constraints, mechanisms, and outputs need to be comprehensively considered [1], which poses great difficulties for wind turbine state modeling.Some new methods/developments in this field also take into account practical operational constraints based on mechanism modeling [2].Besides, intelligent diagnosis methods, such as the expert system and the artificial neural network (ANN), are widely used in fault diagnosis of rotating machinery [3,4], but there are still some defects in practical application.For example, the expert system is difficult to establish a knowledge base and verify its completeness, its fault tolerance is poor, and there are no effective methods to identify error information yet [5].Correspondingly, ANN is comprehensively applied to construct a multi-information fusion device state evaluation model, effectively improving the timeliness and accuracy of state evaluation [6].However, ANN models are limited by the sample selection range and sample time interval, resulting in inaccurate state evaluation results [7].
In this paper, on the basis of training a multivariate state prediction model with historical operating data of wind turbines, this article used two state evaluation methods based on adjustable smoothing parameters and sliding window similarity to establish wind turbine state evaluation models.Through verification with historical wind turbine state data, the effectiveness of the two methods was compared and proven.

Method introduction
The evaluation of wind turbine status was based on the multivariate state prediction model.On this basis, two state evaluation methods were used to evaluate the predicted data, and the results of the two state evaluations were verified based on actual fault data.

Multivariate state prediction model
Multivariate state prediction technology is a non-parametric and nonlinear modeling method, as well as one of the advanced pattern recognition technologies.It was first developed by the Argonne National Laboratory in the United States and was widely used for detecting abnormal operating parameters of equipment.For multivariate state prediction modeling, the construction of a process memory matrix was the most crucial step.The algorithm flow for constructing memory matrices in this article is shown in Figure 1: where D is the process memory matrix, L is the key variable, n0 is the number of equal intervals for L, i L is the average value of the key variable in each interval, Ki is the number of state vectors in each key variable interval, G is a decimal point, and M is the total number of historical state vectors.

State evaluation based on adjustable smoothing parameters
Taking the changes in working conditions as the horizontal axis and the triggering rate of warnings as the vertical axis, the logical relationship should be pure Boolean logic.Therefore, small oscillations of residuals near the threshold will lead to frequent changes and significant differences in warning results.To smooth the triggering rate of warning, a method of adjusting the warning threshold with adjustable smoothing parameters was proposed.This method drew on the basic trust allocation function theory in the Dempster model, which was used for uncertainty information processing.Adjustable smoothing parameters also existed in the fusion target recognition process of basic trust allocation functions.When the state judgment logic was Boolean logic, which was described by solid lines in Figure 2, the basic trust allocation function could generally be in the form of dashed lines in Figure 2. In order to smooth the output of the judgment state, adjustable smoothing parameters were introduced.A function model was established by adjusting the threshold of the adjustable smoothing parameters with the corresponding parameters of different working conditions.The threshold function could be derived from the basic trust allocation function and expressed as Formula (1): where i J and i h are adjustable smoothing parameters and adjusted thresholds respectively; a is the confidence allocation coefficient; i r locates the parameter range by using the average values of different variable parameters under normal operating conditions; m is the ratio of actual output to rated output, used to represent different working conditions.

State evaluation based on sliding window similarity
The similarity between device status and normal status is used as a criterion for evaluating device status, and the more similar it is to normal status, the healthier the device status is.So the evaluation function of similarity utilizes the Euclidean distance E, which is most commonly used in industry, to comprehensively evaluate the state differences in multi-dimensional vector space.Its formula is shown as Formula (2): The Euclidean distance is standardized to the range of 0-1 in practical applications, thus obtaining the fan state evaluation function as shown in Formula (3): where the vectors 1 X and 2 X are observation vectors and estimation vectors respectively, and k is the order of m variables.However, during the operation of the wind turbine, variables of different importance contain different amounts of information, and small changes in some variables can cause significant equipment failure hazards.At the same time, the reliability of variable measurement also needs to be considered.Therefore, it is necessary to assign a state evaluation contribution to different variables, and the similarity could be further defined as Formula (4): In the formula, k O represents the weight of the k-th variable on the impact of faults, which is related to the amount of fault information contained in each variable and the reliability of data measurement.

Case study
The case study was based on actual operating wind power plant data, with a sampling time of 1 minute, to sample statistical fault record data.The input of the state evaluation model included 24 variables, as shown in Table 1.By using two state evaluation methods to evaluate and compare the output of the multivariate state estimation model of wind turbine power generation equipment, it could be concluded in Figure 3 that the sliding window similarity state evaluation method (N=8) had a 9-minute earlier warning time compared to the adjustable smooth parameter state evaluation method, which was faster.In the process of state evaluation, it was found that the sliding window parameters would affect the warning time, so the influence of sliding window parameters on the warning sensitivity was also analyzed and discussed in Table 2. From Table 2, it could be inferred that the threshold for state warning increases slightly with the increase of the sliding window width.Based on the filtering effect of the sliding window, a smaller window width (N=2,4) may cause repeated alarms due to amplitude fluctuations, resulting in unstable state evaluation.However, if the window width exceeds a certain amount (N>8), the width change will not significantly affect the sensitivity of fault alarms, and the alarm time will have a very small delay.

Conclusion
Based on the state prediction model, by comparing the state evaluation methods based on adjustable smoothing parameters and sliding window similarity judgment, it is found that the sliding window similarity judgment method has the optimal warning performance.In addition, the selection of sliding window width needs to be determined based on the actual situation of the wind power generation system.Based on the sliding filtering effect, an increase in window width will cause a slight decrease in the fault warning threshold.If the window width is too small, it will lead to repeated alarms.Exceeding a certain window width will not significantly affect sensitivity, but there will be a slight delay in time.

Figure 1 .
Figure 1.The algorithm flow for constructing memory matrices.

Figure 3 .
Figure 3. Two state evaluation methods for early warning results.

Table 1 .
The input variables of the state evaluation model.

Table 2 .
Comparison of abnormal evaluation sensitivity.