Let's wrap up our search for an energy improvement assessment methodology that is accurate, cost affective and reproducible. So far we've thrown out nacelle power curves due to low accuracy, conventional power curve tests because they are expensive and hard to reproduce.
Now we've landed on side-by-side analysis, where we use unmodified control turbines to measure the performance of the upgraded test turbines. But, we've identified deficiencies in the typical binning approaches used for side-by-side analysis, due to their unacceptably high uncertainty. Especially when we want to try to account for how the test to control turbine power relationship changes with other variables such as wind direction.
Luckily, these days there is a rich eco system of data science tools to enable machine learning, allowing us to model more sophisticated relationships between signals in the wind plant data. Using machine learning, we can significantly reduce the uncertainties compared with conventional binning approaches, sometimes even cutting the uncertainty in half and greatly expanding the lower bounds of performance improvements we are able to measure. And, since these models are typically freely and openly available, they're also cost affective and reproducible. This makes them the ideal tool for assessing energy improvement from wind turbine upgrades.
Even though these models can be powerful, if used incorrectly they will give incorrect results. This is why you need a team with both wind domain and machine learning expertise to apply them properly.