Skip to content
HourProof

Agent

AI agent

How the agent moves flexible load and buys hourly certificates, tested walk-forward on real hours.

In plain terms: An AI agent decides a day ahead when to run flexible work and which extra hourly certificates to buy, so more hours run on clean energy.

What the agent does

Every day, before the day starts, the agent makes two decisions for the next day. It is tested on real hours it never trained on.

  1. Forecast

    It forecasts tomorrow's hourly wind production and the buyer's load, using only data published by the decision time.

  2. Move flexible load

    It moves part of the into hours with more wind.

  3. Buy hours

    It buys extra certificates for the hours its own portfolio will not cover, within a fixed budget.

Method note from the backtest: Flexible share and top-up budget are declared scenario parameters; every MWh value is real Energinet data; decisions use only information available at D-1 17:30 UTC.

Hourly coverage by strategy

The share of the buyer's use covered by certificates from the same hour, over the test half-year. Higher is better.

Test period Jul–Dec 2025

On the second half of the year

Walk-forward on real hours, 2025-07-01 to 2025-12-30. Flexible share 10% of each hour's load (scenario parameter); extra certificates 4.8% of load, and the comparison portfolio gets exactly the same extra MWh.

  • Fixed portfolio70.5%
  • Bigger portfolio, same MWh71.7%
  • Agent: moving flexible load71.4%
  • Agent: moving load and buying hours75.7%
  • 76.7%

Bars start at zero. Perfect foresight uses the same policy and budget fed with the actual load, wind and availability; no real strategy can reach it.

How good are its forecasts?

Average error per hour on the test half-year; lower is better. The agent's load forecast is less accurate than simply repeating last week's load.

Gain over the same MWh
Test period Jul–Dec 2025
+4.0 points

Agent's hourly coverage minus a bigger fixed portfolio that gets exactly the same extra MWh.

Wind error, published forecast
Test period Jul–Dec 2025
514 MWh/h

Energinet's own day-ahead wind forecast, as a baseline.

Wind error, agent's model
Test period Jul–Dec 2025
382 MWh/h

The agent's own wind model.

Load error, agent vs last week
Test period Jul–Dec 2025
2.3 vs 2.1 MWh/h

The agent's load forecast against simply repeating last week's load.

This run

Its plans, on-chain

Each daily plan is committed as a and revealed afterwards; the contract checks that every revealed plan matches its commitment.

AI agent commits its 7 day plans (certificate purchases, CO2 forecast, planned load shift) on-chain before each day starts

0/7 committed before their day started (chain clock) — on a public chain a past week is necessarily committed late; see `live` mode

Agent reveals its plans; the chain checks them against the commitments

7/7 plans revealed and matched

The run replays a past week, so on a public chain its plans can only be committed after the fact. They still cannot change between commitment and reveal; committing ahead of time is what the live mode is for.