Seasonal naive
Repeat the latest eligible observation at a 168-hour seasonal lag.
A transparent weekly reference; changing weather and holidays can break repetition.
Methods & data
Evidence and coverage
Checking the published snapshot…
Sources & units
Energy is hourly kWh from Elhub. Overview totals use MWh and chart axes may use GWh. Production and consumption sum disjoint base groups. Their difference is not a measure of cross-border flows.
Weather comes from Open-Meteo historical weather, using ERA5-Seamless. Area views use one fixed city proxy per area, not spatial averages. Snow drift uses the selected coordinate. Reanalysis and revised energy records do not reconstruct what was known historically.
Forecast models and evaluation
Forecast model cards
The flagship task predicts 24 hours of household demand in all five areas. Energy becomes eligible after the interval ends plus an assumed 48-hour publication lag; historical weather uses 120 hours. These are experimental assumptions, not verified publication archives.
Ridge
Regularized regression with calendar, eligible demand lags, rolling summaries and historical weather.
Fold-local imputation and scaling; linear relationships can miss interactions.
Gradient boosting
Trees using the same information available at issue time; validation selects the configuration.
Nonlinear associations, not causal weather effects. Performance outside the training range is uncertain.
SARIMAX
State-space forecasting with configurable orders and optional projected weather. The benchmark fixes ARMA(1,1).
Convergence failures are reported. Custom nominal 95% intervals omit future-weather uncertainty.
Evaluation protocol
Parameters are selected on chronological validation dates. Residual quantiles are calibrated before the final holdout. Later origins can refit using earlier observations once eligible, while selection and calibration remain frozen. Failed fits are recorded; comparison uses the same successful origins for every model.
The broader reliability study evaluates 46 scheduled dates across 2025, after validation in 2023 and interval calibration in 2024. These previously reviewed observations support exploratory comparisons, not untouched confirmation. Its uncertainty ranges resample blocks of forecast dates with all regions kept together; seasonal comparisons and results from few blocks remain tentative.
Reading the scores and intervals
MAE and RMSE measure held-out error. MASE uses training seasonal changes as its denominator; MASE below one does not prove a model beat the held-out baseline. Benchmark intervals use empirical 10th/50th/90th residual quantiles. Evaluate coverage, width and pinball loss together. Realized future weather is available only as a separately labeled upper-bound experiment.
Demand sensitivity
Weather-adjusted demand sensitivity
The Demand sensitivitystudy is a saved retrospective analysis of hourly household demand in NO1–NO5. It compares a calendar-only reference, a linear temperature response and a flexible temperature response on later observations. The page shows the study’s fixed training, validation and test dates; changing the dates in other dashboard views does not refit this study.
The temperature curve is an adjusted association after accounting for hour, weekday, holidays and seasonality. Its shaded 95% pointwise band describes uncertainty in the estimated mean response, not the range for an individual hour. Temperature support and held-out error should be read alongside the curve, especially near the edges of the observed range.
Each area uses one fixed city weather proxy rather than an area-wide temperature measure. The result is neither a causal temperature effect nor an operational weather forecast; demand scales differ across areas, so raw error values should not be used as a ranking without that context.
Validation case studies
Recorded findings
Recorded examples from the September 2026 validation snapshots, independent of current filters and synthetic fixtures.
Did Eastern Norway generate as much as it used?
19.72 vs 32.42 TWh
Production / consumption · complete coverage.
Compare the year’s daily curvesEvidence
Totals sum disjoint base groups over complete UTC days. The balance does not measure imports or exports.
Read the recorded numerical checkCan a model improve on repeating last week?
29.96% lower MAE
Gradient boosting’s hourly demand error was 61,991.89 kWh, compared with 88,508.58 kWh for the weekly baseline.
This small holdout does not establish broad superiority. Its nominal 80% intervals covered only 61.6% of outcomes.
Inspect errors and interval coverageEvidence
Models share the same successful origins. Ridge’s MAE was 1.71% worse than the baseline.
Read the frozen evaluation protocolWhat can a snow-transport model tell us?
44.7 tonnes per metre
The recorded Tabler calculation estimated snowfall-controlled transport at 60.3913° N, 5.3221° E and an indicative 2.13 m Wyoming fence height.
A weather-point model illustration, not a site-specific engineering design. Location and model assumptions affect the result.
Inspect the season and assumptionsEvidence
The recorded run used 3,000 m transport distance, 30,000 m fetch and a relocation coefficient of 0.5.
Read the point-model validationArchitecture and reproducibility
Architecture and reproducibility
Public sources → validated hourly SQLite and weather snapshots → FastAPI analysis and stored forecast artifacts → Next.js charts and tables. Refreshes publish atomically; ordinary area views read local snapshots. Only explicit custom jobs fit forecasting models, with one worker, a bounded queue, timeout and cancellation.
UTC intervals use an exclusive end; browser date controls include both selected dates. Calendar features and daily profiles use Europe/Oslo. Profiles exclude partial, incomplete and 23/25-hour daylight-saving days from their 24-hour comparison, while retaining those days for inspection. Missing hours remain on the regular grid. Aggregation and display sampling are disclosed alongside each analysis.
Exports carry source versions, parameters and units. Forecast artifacts also record training windows, issue times, code fingerprints, dependencies, failed folds and calibration. Reproducing an exact result requires the same source snapshots and environment.
This independent continuation of the original IND320 project preserves its notebooks, source attribution and historical validation in the repository. See its validation reports for full methods and recorded checks.
Local experiments
Custom forecast jobs are disabled on the public deployment. A local installation can run bounded evaluations and SARIMAX jobs with progress reporting and cancellation.