Eight Years of Yaw Data: What 3,118 Days Across 104 Turbines Taught Us

We recently closed out a long-running yaw stability and vibration study covering an entire 104-turbine fleet from 2018 to 2026 — 3,118 days of records. This write-up walks through the method: how we cleaned the SCADA record, defined stability thresholds at multiple levels, separated genuine yaw misalignment from wind-driven noise, and what the fleet-level patterns mean for maintenance planning.

Why yaw stability deserves this much attention

A turbine that tracks the wind poorly loses energy every single day, and it wears itself out doing it. Persistent yaw misalignment loads the blades asymmetrically, shows up as measurable vibration in the nacelle and drivetrain, and accelerates wear on the yaw drive and braking system. The problem is that "misaligned" is not a binary state — every turbine spends some time off-optimal while the control system catches up. The question that matters for maintenance planning is which turbines are persistently off, by how much, and whether that behaviour is getting worse.

Cleaning eight years of SCADA data

Before any analysis, the record has to be trustworthy. Eight years of 10-minute SCADA averages from 104 turbines is roughly 40 million rows, and a meaningful slice of them are wrong: frozen sensors repeating the last value, clock drift between data loggers, unit mismatches after controller updates, and long outage gaps that must be excluded rather than averaged over.

Our cleaning pass removed downtime periods, flagged flat-lined channels, cross-checked wind speed against nacelle direction and power output for physical consistency, and kept only days where the full sensor set was healthy. What survived was 3,118 days of usable, comparable records per turbine — the base for everything that follows.

Defining stability thresholds — at more than one level

A single threshold (say, "misalignment above 5° is bad") hides more than it reveals. We characterised each turbine at three levels:

Normal band — momentary deviation within the control system's normal correction cycle. Every healthy turbine lives here part of the time.
Watch band — deviation that persists across multiple averaging windows in moderate winds. Not yet a defect, but a fingerprint worth tracking.
Action band — sustained deviation, or a watch-band signature that repeats across wind sectors. This is where an inspection or a caliper adjustment earns its place in the schedule.

The same tiering was applied to vibration: RMS levels in the frequency bands tied to yaw activity, compared per wind-speed bin, so a turbine working hard in strong wind is not penalised for the weather.

Separating misalignment from the wind

This is where most naive analyses fail. A turbine sitting 4° off the wind in a 3 m/s drift is a different animal from one sitting 4° off in a 14 m/s gust front — the first is likely the control system hunting, the second is likely something mechanical. We correlated deviation against wind speed, wind direction sector and rate of change, filtered out ramp events and direction transients, and looked only at quasi-steady operating points. What remains is behaviour that belongs to the turbine, not to the weather that day.

What the fleet actually showed

Across the fleet, most turbines sat comfortably in the normal band the large majority of the time. But a small, well-defined cluster showed persistent watch- and action-band signatures — concentrated enough to be systematic, not random. Their vibration signatures pointed at the yaw braking system: the classic pattern of calipers that no longer hold the nacelle cleanly against wind-load reversals. That cluster became the target list for a yaw caliper replacement and adjustment campaign, with before/after vibration verification built into the scope from day one.

From analysis to maintenance planning

The deliverable was never a report for a shelf. It was a ranked work list: which turbines, which intervention, what evidence supports it, and what the verification measurement should show afterwards. Planned into a weather-buffered schedule, the corrective work could be sequenced against wind-speed history rather than optimism — and the follow-up data closes the loop on whether the intervention worked.

Takeaways

Clean the record first — conclusions built on frozen sensors and outage gaps are worse than no conclusions.
Use tiered thresholds, not one line in the sand: persistence and repetition matter more than single exceedances.
Correlate with wind conditions before calling anything a defect — the weather writes a lot of false positives.
Fleet-level patterns turn individual oddities into a targeted, justifiable campaign.
Build verification into the scope from the start, so the campaign proves itself in data.