In order to achieve their desired production and revenue goals, wind farm operators must understand and manage the risk associated with the degraded performance of wind farms. The acquisition and storage of data are fundamental in gaining insight into whether the maximum available power is being achieved and where a shortfall may be occurring. By answering “what” and “where,” wind farm engineers can better understand the risk associated with achieving production goals and can recommend remedial actions to address the shortfall.
There are three categories of data analysis that form a natural workflow on building insight and providing enhanced decision-making support: descriptive analysis, post-event diagnostics and prognostics.
Descriptive analysis. This category of analysis identifies the main features of a data set through established statistical calculations and visual representations. Examples of descriptive analysis include calculations, such as mean and variance; and visualization techniques, such as a scatter plots, histograms and box plots.
Post-event diagnostics. This analysis seeks to identify the cause and effect of system change. In this context, system change is generally an abrupt change that will exceed a predefined threshold.
Prognostics. This method of analysis seeks to predict system change. The state of the system is monitored for a subtle change relative to a reference behavior, where the change is deemed to be trending toward some undesirable condition.
These definitions are explored by considering the relationship between power output and the key inputs that influence power output for a simulated wind farm containing Type III wind turbines. The presented example is illustrative only.
The main features of the data gathered from system operation are exposed through standard statistical computations and visualization methods. The primary goals of descriptive statistics are to inform and form the core of subsequent analysis. The larger the data set is, the more informative and accurate the descriptive analysis becomes. Therefore, the longer a wind farm operates, the more data it will provide.
The first relationship to be considered is that between wind speed and power output, which generally takes a form similar to that shown in Figure 1.
Note that there is not a one-to-one relationship between input and output, but rather for a given wind speed, there is a spread of possible power outputs, from which mean and standard deviation can be calculated.
Also shown in Figure 1 is a convex hull, which is a convenient and rapid computation used to define the boundary of the surface.
Because many wind farms have multiple wind turbines with similar specifications, the output of descriptive analytics forms a strong basis for informing post-event diagnostic and prognostic analysis. A comparison of the spread of multiple turbines containing similar specifications can reveal the turbine with the most efficient boundary condition.
Figure 2 shows an example of two boundaries taken from two turbines with similar specifications. The boundaries are offset, with Turbine 1 exhibiting a decrease in power output compared to Turbine 2 for a given wind speed. Turbine 1 is, therefore, less efficient than Turbine 2.
Figure 3 shows a surface plot of wind speed versus power output, where power output has dropped outside an expected boundary condition. This indicates that some event has occurred that is detrimental to the expected performance. Because Figure 3 yields no insight into the cause of the reduction in power, further inputs must be considered. In this case, standard diagnostics would be applied, and the inputs to the system would be examined pre-event and post-event.
Figure 4 shows the response of a wind turbine that is operating exclusively within the boundary layers. Based on this data, there is no indication of whether the turbine is operating in a degraded state, as a monitoring system that relies on violation of the boundary will not register an alarm. In scenarios such as this one, the development of a predictive model becomes an especially valuable tool to assist in the detection of degraded performance.
The heart of prognostic analysis is predictive modeling. Although several different modeling paradigms exist – such as ARMAX, regression trees and neural networks – they all rely on being constructed from a data set that represents the desired operation and that has sufficient inputs and outputs to form an accurate model.
The selection of data that is representative of the desired operation performance can be based on the computation of boundaries, as shown in Figure 2. A boundary can be constructed for each of the wind turbines with similar specifications, and the wind turbine with the boundary of lowest surface area can be selected as the reference.
Several inputs can be considered – such as wind speed, blade pitch, electromagnetic torque reference and generator shaft speed – in order to develop a model that predicts power output. In order to select the most appropriate model architecture, a number of models that use different combinations of the inputs should be developed. The models can then be tested on additional data sets to determine model robustness in identifying performance degradation.
Figure 5 shows an example of mean squared error (MSE) between measured data and the output of two predictive models: a regression tree and a neural network. In this case, turbines 20 through 30 were degraded in comparison to turbines 10 through 20, and turbines 30 through 40 were degraded by a larger margin. In this case, the degradation was due to an increase in shaft friction.
The error between the models and the measured outputs can be used to indicate the level of performance degradation. The engineering team can then use this information to help decide what appropriate action can be taken, such as bringing forward a scheduled maintenance.
By applying predictive analytics techniques, the engineering team can detect degraded performance earlier and, in turn, make a timely and informed decision on the action necessary to address the degradation. w
Industry At Large: Wind Farm Operations & Maintenance
Big Data Means Big Insight For Wind Farm Operations
By Graham Dudgeon
The management of risk associated with degraded wind farm performance is critical to achieving desired production and revenue goals.
NAW_body NAW_body_bi NAW_body_b_i NAW_body_bNAW_body_i