Currently, wind turbine control algorithms are designed based on computer models of air flow over the turbine blades. These models have to make assumptions about the environment in which the turbine will be installed such as site topography, ground roughness, altitude, and weather conditions. However, in practice every turbine is installed in its own micro-environment which is unlike any other turbine installed anywhere else in the world. In addition, the computer models assume an ideal turbine with the exact designed geometry and surface smoothness of the three blades. In the real world, blade geometry can have small variability and surface smoothness can degrade slightly over time. Until now, no performance characterization tools existed which were accurate enough to reliably detect these subtle issues on individual turbines. This creates a challenge for in-service turbine performance optimization.
This challenge is why our team went to work to develop statistical methods to not only enable detection of subtle turbine inefficiencies, but also to facilitate automatic correction through control algorithm optimization. The optimization system works by combining Bayesian statistical analysis with experimental design methods to characterize turbine performance on a whole surface throughout a space of control parameter settings. Then the system uses this surface to find control strategies which perform optimally, better than the default settings.
We have developed the system to be automatic and remotely executed. From our central operations center, turbines anywhere in the world can be commanded to execute the optimization routine. It is pretty incredible with you think about it. From a single command station, every turbine at a wind farm can be tested simultaneously and each will receive optimized parameter settings unique to its own micro-environment and hardware condition.
As an example, Figure 1 shows the result of applying the optimization algorithm to a turbine undergoing alternate blade pitch strategies. The three blades on a wind turbine are continuously pitched toward the wind or away from the wind to maximize energy capture while keeping physical loads within tolerance. Each point on the plot gives the measured power in KW vs. wind speed in m/s during a 10 minute time interval during the experiment. The black data were taken under the default control algorithm and the red data were from an optimized algorithm, and the curves were fit using the Bayesian statistical analysis. A subtle improvement can be seen in the range 8-12 m/s wind speed. However, due to the sensitivity of the statistical analysis we can detect this improvement with high confidence, as shown in Figure 2.This type of updated control strategy is offered as an aftermarket solution to the installed base, which increases annual energy by 2% on average or up to 4% in seasonal power. Scaling this improvement up to a whole wind farm can strongly impact a farm operator’s bottom line while making the most of this precious energy resource.
Figure 1: Raw data and fitted power curves under two control algorithm conditions. Condition 2 is an optimized blade pitch control algorithm.
Figure 2: Estimated power difference, optimized setting minus baseline setting with 95% uncertainty bounds.