Day-ahead wind generation forecast that beats persistence

Nordwind sells electricity from 1.4 GW of onshore wind parks in northern Germany on the day-ahead market. Every day before noon it has to bid 24 hourly volumes for the next day. Whatever it gets wrong is settled at the imbalance price, which last year averaged about €45 per MWh of error on top of the lost sale. Nordwind currently buys forecasts from a vendor but wants an in-house model as a check and a fallback.

The forecasting lead has set a clear test. Use only data that would have been available at 11:00 on the day before delivery: - hourly wind generation for one German control area from SMARD, the Bundesnetzagentur market data portal, scaled to a per-unit capacity factor; - weather forecasts for 6–10 locations in that area (wind speed at 10 m and 100 m, direction, temperature, pressure) from an open source such as Open-Meteo's historical forecast data or DWD open data.

You will build two baselines, persistence (tomorrow equals today at the same hour) and climatology (the average for that month and hour). Then you will train a gradient-boosted model and a small neural network on the weather features. The test is the full most recent year, held out. The target is a normalised RMSE at least 40% below persistence.

Done looks like a reproducible pipeline that builds a forecast for any date, a results report with error by lead hour and by season, a short comparison of the two models, and an estimate of how many euros of imbalance cost a year the better model would save on 1.4 GW.