Carbon
Carbon-aware load
Shifting flexible load toward low-carbon hours, and proving residual emissions in zero knowledge.
In plain terms: The agent can also move flexible work to hours when the grid is cleaner. The buyer then proves its remaining emissions without revealing its usage.
Move load to cleaner hours, then prove the result
A learned day-ahead forecast of Denmark's hourly grid CO2 intensity moves 10% of each hour's load (a declared scenario parameter) into cleaner hours of the same day. Test 2025-07-01 to 2025-12-30, run once with choices frozen on validation. Forecast error 20.3 g/kWh vs 38.4 for last week's value (47% lower).
The buyer then proves its with a , using the intensities posted on-chain and the same commitment to its load as the coverage proof.
Emissions by strategy
Lower is better.
The buyer's location-based emissions, by strategy
4,392 hours of the IT sector's load. Bars start at zero; the differences are in the right-hand column.
- No shift48,840 t
- Shift for MWh matching47,889 t−1.95%
- Carbon-aware, last week's intensity48,546 t−0.60%
- Carbon-aware, learned forecast47,634 t−2.47%
- Carbon-aware, actual intensity (reference)46,856 t−4.06%
Emissions use Energinet's hourly factors. The reference row knows the actual intensity in advance; no real forecast can.
The learned forecast cuts more emissions than using last week's intensity
Test period Jul–Dec 2025913 t less
95 % interval 695–1,162 t; above zero in 100% of resamples.
The learned forecast cuts more emissions than shifting load for MWh matching
Test period Jul–Dec 2025255 t less
95 % interval 106–400 t; above zero in 100% of resamples.
The run's week, hour by hour
How clean the grid was each hour, what the agent expected the day before, and where it moved flexible load.
Hourly grid intensity: the agent's forecast and the actual
The actual values are Energinet's location-based intensities, as posted on-chain. The forecast was made the day before.
Week of 2025-11-10 (UTC). Forecast error this week 15.6 g/kWh (last week's value: 40.5). Source: demo run 20261017T141639Z (done).
The week, shaded by intensity
Darker squares are dirtier hours. A dot marks an hour into which the agent planned to move flexible load.
Hover over or focus an hour to see its numbers.
Week of 2025-11-10 (UTC). Shading runs along the scale, from the cleanest hour of the week to the dirtiest. Forecast error this week 15.6 g/kWh (last week's value: 40.5). Source: demo run 20261017T141639Z (done).
Caveats
Test period Jul–Dec 2025
- Average, location-based grid emission factors (Energinet's declaration, 125 % method) for a buyer's accounting; not marginal emissions and not physically avoided emissions.
- The 10 % flexible share (and the +25 % hourly cap) is a declared scenario parameter, not a measured property of the IT sector's load.
- One test half-year (Jul-Dec 2025); no claim that the results generalise beyond it.
- Training uses Energinet's Final declaration as if it were available at decision time; Final is published later (live mode uses the Initial version for the lag features).
- Unmatched hours are valued at the hourly grid average intensity, not a residual-mix factor.
- Nothing here is audited.