Solar.

The median relative imbalance of a solar plant per quarter-hour is roughly 2.4 times larger under cloudy or unstable skies than under clear skies. Measured on per-settlement-unit imbalances in Portugal, it goes from about 11% to about 26%.

There is a trap in this reading: on daily sums the effect inverts, because level and internal netting hide the variation. It only shows when measured per quarter-hour, normalised by the schedule. Anyone who does not make that separation calibrates the model on the wrong day.

A second term is even less visible. Between the forecast the day-ahead bid was made with and the current forecast there is a drift — and that drift is observable before intraday gate closure.

The solar resource model is conditioned on sky regime, expected price and horizon, before any market decision. Under clear skies the distribution tightens; under unstable skies it widens. The decision sees the right distribution in each case.

• A distribution, not a point. What we deliver is a distribution per quarter-hour. A central value alone is not enough to decide under asymmetric cost.
• Physical parameterisation of the plant. Connection capacity, DC/AC ratio with losses, and resource profile by curve — fixed tilt or tracker. Clipping shows up as a midday plateau and is modelled as such.
• Drift between vintages. The difference between the day-ahead forecast and the current one is treated as a signal in its own right, not as noise.
• Fitted to the plant’s own history. The measurement-to-output conversion comes from the plant’s own series, not from a catalogue curve. It recalibrates monthly, using only setpoint-free hours.