# Fixed seasonal schedules and Great Britain electricity carbon intensity

## Question and scope

Does a three-hour daily electricity-use schedule, chosen separately for four seasons using only 2023 national data, retain lower average carbon intensity than a midnight schedule in 2024 and 2025? The primary comparison is 2025 seasonal versus 00:00 UTC. The 2024 comparison is a prespecified temporal transport check. Both test years remain excluded from all schedule selection.

This is a retrospective accounting benchmark for one unit of electricity distributed uniformly over three hours each day. It measures average generation-related CO2 intensity assigned to consumption, not marginal emissions avoided by changing load. The observed grid remains fixed. No forecast performance, consumer feasibility, costs, storage losses, power-system response, or intervention effect is claimed.

## Data and timing

Use the official NESO national Carbon Intensity API, GET /intensity/{from}/{to}, to obtain half-hourly intervals starting 2023-01-01T00:00Z through 2025-12-31T23:30Z. Retrieve intervals in chunks no longer than the documented fourteen-day limit. Preserve raw responses compressed losslessly, source URLs, retrieval times, and SHA256 digests. Analysis uses intensity.actual exclusively, which the provider describes as estimated intensity from metered generation, in gCO2/kWh. Forecasts and index labels are not outcomes or predictors.

At registration no historical API data for these years have been downloaded or inspected, and no schedule or effect estimate has been calculated. Documentation and literature have been read. General knowledge of diurnal and seasonal variation informed the question. The plan, rather than completed code, is registered before source retrieval.

## Eligibility and validation

Parse timezone-aware ISO datetimes structurally. Filter by interval start, including the start and excluding the end of the study range. Require thirty-minute intervals aligned to the UTC half-hour. Identical overlap rows at request boundaries can be deduplicated; contradictory duplicates cause failure. Require a finite nonnegative numeric actual value and all forty-eight distinct half-hours for a complete UTC day. Do not interpolate or substitute a forecast for missing actual values. All schedules for a comparison use exactly the same complete days. Record every missing or invalid day and its reason, by year and month. If fewer than 95% of days in any year are complete, describe a coverage failure rather than generalizing to the full year. Require at least sixty complete training days in each season.

## Schedules

Candidate starts are half-hour boundaries from 00:00 through 21:00 UTC, inclusive, so all six intervals of a three-hour window lie within one UTC day. There are forty-three candidates. A window's daily cost is the arithmetic mean of its six actual intensities. There is equal electricity use per complete day and equal energy in each interval, not national-demand weighting.

Choose the candidate with lowest mean daily cost across eligible 2023 days for each season: DJF (December, January, February), MAM, JJA, and SON. Thus DJF pools January, February, and December of the same training calendar year. Break exact ties by earliest start. Also choose one global start from all 2023 days, as a secondary comparator. Freeze all choices before evaluating 2024 or 2025. For each test day use its season's training choice.

Prespecified comparison schedules: midnight 00:00-03:00 UTC; evening 18:00-21:00 UTC; the global 2023 start; and the season-specific 2023 start. Also compute the hindsight-optimal candidate separately each test day as an unattainable information benchmark, never a deployable forecast. UTC is deliberate: this study does not implement fixed British civil time through daylight-saving changes.

## Outcomes

For each test year report the mean assigned intensity of each schedule and the paired absolute difference and percent reduction for seasonal versus midnight: 100*(mean midnight - mean seasonal)/mean midnight. Also report the fraction of complete days when seasonal is strictly better, equal, and worse. Report the same differences by all twelve months, plus the fraction of the midnight-to-hindsight gap captured by the seasonal schedule, if that denominator is positive. Report annual global-versus-midnight and seasonal-versus-global contrasts, and evening comparisons as secondary results. Include every planned result regardless of sign.

These are finite-period descriptive quantities for the observed complete days. No p-value or nominal confidence interval will be used to generalize them to other years. The full observed monthly range and complete-day coverage convey temporal heterogeneity and missingness. API estimation/revision error is unquantified. A positive annual difference alone will not imply all months or days improve. No refitting on either test year, alternate training window, or selection of a more favorable baseline is allowed in the primary analysis.

## Checks and reporting

Independently recompute selected windows and annual totals from raw responses using direct interval selection separate from the rolling-window implementation. Use synthetic fixtures for constant intensities and exact ties, a known minimum window, half-hour boundaries, missing values and incomplete days, contradictory duplicates, leap-day inclusion, and training-only schedule selection. Perturb test values and verify learned schedules are unchanged. Verify the hindsight cost is no greater than any fixed candidate every day, annual means match a direct half-hour weighted calculation, and all percentage denominators are explicit.

Publish original code, pinned environment, raw public aggregate responses under their CC BY 4.0 license with attribution, source hashes, the registered plan, exclusions, selected schedules, and all planned outcomes. Declare deviations. Limit the claim to the frozen historical series and methods. If the result only repeats existing work without useful new information, retain the attempt and do not publish filler.