Wind.

Wind forecast error is larger in amplitude and far more persistent than solar. A phase error on a ramp does not wash out in the next quarter-hour: it holds for hours, in the same direction, and accumulates.

Converting measurement into output is also harder. Wind measured at the nacelle is biased by the turbine itself; the turbines in a farm do not see the same wind; and wake effects between turbines change with direction. A catalogue power curve solves none of these three things.

The forecast integrates several public atmospheric models as an ensemble. We do not run our own atmospheric models — we integrate the ones that exist, and the advantage lies in what is done with them downstream.

• Model disagreement as signal. When models diverge, the distribution widens. Disagreement is information about uncertainty, not a problem to be averaged away.
• Physical curve fitted to the farm. The wind-to-power relationship is estimated on the farm’s own series, absorbing nacelle bias, turbine-to-turbine spread and wake effects.
• Ramps treated explicitly. Phase error on a ramp is the largest contributor to wind imbalance and is modelled as such.
• Recalibration with more care. Wind requires longer windows and stricter minimum-sample rules than solar before the coefficient is touched.