Wind-farm imbalance risk: where the money isn't

Hill of Towie is a 48.3 MW wind farm in Moray, Scotland. It sells electricity a day ahead: by 09:20 each morning it commits a bid for every hour of tomorrow, and any deviation from that bid settles at the half-hourly System Price. The goal here is not to build a better forecast — it is to understand the imbalance cost the current bid is exposed to, and decide what, if anything, is worth doing about it.

The brief's cost is (actual − bid) × SP, which nets to about −£32k/yr — nearly free. But that mixes the energy value of a deviation with the cost of being wrong: when the farm over-produces it gets paid for real electricity, which books as profit and hides the fact that energy was sold below the day-ahead reference. Benchmarking against a perfect-foresight bid instead isolates the cost of being wrong — regret:

regret = (actual − bid) × (IMRP − SP) = volume_error × price_spread

This factorises into a volume-error leg and a price-spread leg — and all the structure lives in how they interact. Net regret is +£103k/yr. It is the lens the rest of the analysis uses; the brief's cashflow is reported alongside, never replaced.

Every net in this dataset is a thin residual of large opposing flows, so any aggregate misleads until decomposed by sign and price regime. Split it four ways and the story inverts: the biggest component is a forecast artifact, the most dramatic days are dispatch fiction, and the genuinely hedgeable core is small. The decomposition depends on where you draw the thresholds — so move them, and watch what stays true.

96% of net regret sits in the extreme price deciles — it is a tail problem. And it is directional: in the dearest price decile the farm is short 64% of the time, cleanly clear of the 50% no-bias line. The same calm weather that starves this one farm starves GB's whole wind fleet, tightens the national system, and spikes the price — so the farm is short exactly when buying back is most expensive. The risk lives in Cov(error, price), not in the size of the error.

The most extreme dear-tail days looked like storms that knocked the farm out. Turbine telemetry says otherwise. The dispatch setpoint — the maximum each turbine is permitted to produce — sits pinned at zero for fifteen straight hours while the wind blows a healthy 12–16 m/s. The farm wasn't broken or becalmed; it was commanded off: grid curtailment behind a constrained Scottish transmission boundary, corroborated by national constraint costs running 2.6× their median on those days.

This is the pivot of the whole analysis. Curtailment is dispatch, not a forecast error — no bid can hedge it. And it hits both price tails: fake losses when commanded off into expensive prices, equal-and-opposite fake gains when commanded off into cheap or negative prices. Its P&L nets to fiction, so it is excluded entirely — never netted, never defended as a number.

Can the genuine core be flagged a day ahead? Using only leakage-safe signals known by gate time — forecast temperature for icing, ECMWF–GFS disagreement, national system tightness — a cumulative gains curve gives AUC ≈ 0.66, and the top decile sits below random. The signals identify a risky class of day, but not which days will carry the loss; the biggest events are forecast busts and sub-3h ramps, invisible day-ahead.

Mitigation is sized on an exact revenue identity — ΔRev = (bid_new − bid_old)(IMRP − SP) — fit on 2023, evaluated out-of-sample on 2024. A standing compression de-bias (no forecasting needed) recovers only ~£16k on 2024, CI [−£7k, +£40k]: raising the bid earns +£153k in cheap periods and gives back −£137k in expensive ones. Conditioning on a day-ahead price signal made it worse. The flag-shade is marginal.

The same test — split gross from net, split by regime, check concentration — overturned three of my own headline numbers before the decomposition stabilised. That is not flailing; it is one instrument applied at each level, each time revealing structure the aggregate hid. Stated plainly:

• — Every net is a thin residual of large opposing flows, so any single aggregate misleads until decomposed.
• — The hedgeable core is a range (£10–80k), concentrated in ~10 days/yr — so any savings figure has wide bands.
• — Curtailment compensation is unprovable from this data (no unit-level balancing-mechanism series); it's dispatch either way, so not hedgeable, but the exact settlement is a named data gap.
• — Two years, one site, a simplified settlement model — directionally solid, not a P&L guarantee.