wind-up v1

[aclerc]

v0 is a single-method tool: it measures turbine-upgrade uplift with a binned power-curve, test-vs-reference method. v1 turns wind-up into a platform in which alternative uplift methods are pluggable and objectively benchmarked on synthetic datasets with known ground truth. The driving goals are accurate and precise results from short campaigns and richer conditional information about how an upgrade performs (uplift by wind speed, uplift in wakes vs free-stream, day vs night, by direction/stability, etc.).

Metrics:

• Campaign uplift P50 bias and spread
• Conditional uplift distribution: can measure per bin bias and spread of uplift distributions (vs wind speed, wind direction, TI, etc). Uplift vs wind speed is generally the most important for pre-con inputs but it has to be a stable, free-stream wind speed estimate (separate bit of work to construct that, eg invert a warranted power curve). For certain upgrades other conditional results are very important (example wake steering also wants vs direction and atmospheric stability).
• Uncertainty quantification: harder to judge but can verify campaign uplift actual quantiles vs those implied by the reported uncertainty (eg verify that ground truth uplift is below P95 exactly 5% of the time)
• AEP uplift P50 bias and spread
  • updating the pre-construction model is probably the best approach in most cases. This would generally require deltas to the power curve and thrust curve. Industry standard is still a LT distribution of wind speed and direction. So one task is to translate the X-month measured "pre-con inputs" to long-term "pre-con inputs". Then one can run pre-con model for long-term weather. Can validate the pre-con model agrees with campaign results when it is run just for campaign weather as a sanity check or calibration step.
  • In some cases the AEP calculation will be done with a specialist tool (example FLORIS for wake steering). In those cases wind-up needs to report out whatever these tools need. FLORIS modelling would need more complex things like Pp (can be derived from uplift vs steer command).
  • A 'ground truth' AEP can be calculated if the source data has sufficient years. HOT has 5 clean years (2016-2020) so an AEP comparison based on those 5 years can be done.

Design considerations:

• Robustness to northing jumps and other common sensor drifts and jumps (especially anemometry changes)
• Robustness to reference turbine performance changes. Reference turbine performance can change in time which introduces risk of bias. Reference turbine wind speed is even more likely to change in time (eg due to anemometer calibration change)
• Robustness to occasional abnormal operation of the test and reference turbines. Easy for an analyst to identify, decide what is normal (ie will repeat in future) and should be kept and filter out what is not