Projects /
Data & AI · 1st place
Imbalance Price Forecast for Turku Energia
Our team won first place in the Turku Energia challenge at the Since AI 2026 hackathon. We forecast the Finnish electricity imbalance price 10 hours ahead, in 15-minute steps, and built a "situation room" dashboard for energy traders.
Since AI 2026 was a hackathon in Turku on 6–8 November 2026. Our team of five took the challenge from Turku Energia, and we won first place.
The problem
Electricity production and consumption have to match at every moment. When a company’s plan and what actually happens don’t match, the difference is an imbalance, and it’s settled at the imbalance price. In Finland that price is set separately for every 15 minutes, and it can jump far above or below the normal day-ahead (spot) price. Seeing those jumps early means lower costs and better-timed trading.
The task: from every full hour, forecast the next 10 hours (40 quarter-hours):
- the imbalance price, as a median with an 80 % range (p10, p50, p90)
- the probability that the grid needs up-regulation, down-regulation or neither
- the probability of an exceptional price, more than 50 €/MWh above or below spot
The trick
The imbalance price is basically spot, which is already known a day ahead, plus a jump up or down. So we don’t guess the price from scratch. We guess the jump. Just doing “spot minus a little” already beat the organisers’ baseline by about a third. The jump depends mostly on how wrong the wind forecast is, the time of day, and big price changes from one hour to the next.
My part: the data
I was responsible for the data layer. That covered the external sources, the features, and above all the availability layer, the rule that decides whether the whole thing is honest.
Forecasts were scored on a past period, so it’s very easy to cheat by accident: use a number that wasn’t published yet at the time of the forecast, and the model looks great without ever being useful. So every row from every source gets a first_usable_origin_utc column. It’s computed from a conservative publication delay and rounded up to the next full hour, and a forecast made at time t can only use rows where that’s ≤ t.
On top of the organisers’ starter package (Fingrid, ENTSO-E, Nord Pool spot), I wrote fetchers for:
- Fingrid Open Data: real-time 3-minute system data, production plans, wind capacity and mFRR bids
- FMI open data: measured wind and temperature at coastal stations
- Nord Pool UMM: planned and unplanned power plant outages, with their publication times
- Elering, Energinet and Open-Meteo: neighbouring markets and weather
- 15 months of extra history, so the models could train on last autumn and winter instead of only four months
The rest of the team
The models are LightGBM, blended from two independent approaches. Heavier runs went to the LUMI supercomputer. The dashboard shows what’s coming in the next 10 hours, warns about spikes, explains why in plain words with an LLM, and lets you “time travel” to any past hour to see exactly what the forecast said then. When the two models disagree, it shows a “conflicting signals” warning instead of hiding it.