German Grid-Load Forecasting

A daily and fixed 24-hour-ahead German electricity-load forecasting case study built around a practical question: how accurately can grid demand be predicted using only information available when the forecast is issued?

Key findings

• The daily ensemble reached 1.87% MAPE and reduced error by 73% against the previous-day baseline.
• Aligning hourly forecasts with the daily total reduced the selected hourly model's MAE by 4.3%.
• The selected residual hybrid and direct LightGBM were statistically comparable in the final evaluation; the evidence does not establish a clear winner.
• The adaptive 90% interval achieved 89.95% empirical coverage, the closest tested method to its stated level.

The dashboard connects the forecasts, error patterns, uncertainty intervals, and model comparisons in one place.

Explore the dashboard · Read the model card · Read the technical report

Features

• Data — Fetches German net load from SMARD (filter: 410, hourly, data/raw/) and weather from Open-Meteo archive and previous-runs forecast APIs. Uses a day-ahead weather strategy: archived forecasts issued 24 hours before the target day from 2024 onward; earlier rows fall back to the previous day's observation (config.yaml -> features.weather_strategy).
• Features — Calendar features (holidays via the holidays package), Fourier terms (weekly and yearly), HDD/CDD from temperature, and lagged load (lags: [1, 7] daily, [24, 168] hourly).
• Models — Daily models: SARIMAX, XGBoost, LightGBM, RandomForest, and a frozen ensemble of SARIMAX + XGBoost + LightGBM selected on the 2024-2025 validation period. Hourly models: a residual-hybrid forecaster (linear trend plus LightGBM residual model), a direct multi-horizon LightGBM model, and a direct ridge model (src/loadfc/models/).
• Uncertainty quantification — Fixed and sequentially adaptive conformal prediction intervals: symmetric split conformal per horizon, CQR, and adaptive conformal with calibration window and gamma drift (src/loadfc/evaluation/conformal.py).
• Evaluation — Metrics (MAE, RMSE, MAPE, MASE), baselines (naive 1-day, seasonal naive 7-day), day-block bootstrap ablations (rolling origin), Diebold-Mariano tests, error slices, weather ablations, and drift analysis (src/loadfc/evaluation/).
• Reproducibility — Release manifests keep generated tables, the dashboard, and the report tied to the same committed experiment (src/loadfc/tracking.py).
• Report — LaTeX technical report generated from run outputs (report/technical-report-en.tex, compiled with Tectonic).

Quick start

Requires Python >=3.12,<3.14, uv, and Tectonic 0.16.9.

uv sync --extra explain
uv run python scripts/run_pipeline.py --config config.yaml

scripts/run_pipeline.py runs the full chain: fetch data, build features, fit and evaluate models, compute intervals and analyses, compile the LaTeX report, validate the staged bundle, and atomically promote it. Add --refresh-data to re-download raw data or --config <path> to use another config file. The release command requires a clean committed source tree.

Detailed CSV evidence is generated on demand rather than stored in Git. To reproduce it without a local setup, run the manual reproduce release evidence workflow in GitHub Actions. A successful run uploads results/ and the regenerated report PDF as release-evidence-<commit>, retained for 30 days.

To inspect the local static dashboard after regeneration:

npm ci --prefix dashboard
npm --prefix dashboard run build:docs
python -m http.server --directory docs 8000

Visit http://127.0.0.1:8000/.

Project layout

config.yaml                  All run settings: data ranges, split dates, model hyperparameters
src/loadfc/
  data/                      SMARD and Open-Meteo fetching, dataset building, validation
  features/                  Calendar, Fourier, lags, weather features, feature assembly
  models/                    SARIMAX, ML (XGBoost/LightGBM/RF), hybrid and direct hourly models
  evaluation/                Metrics, baselines, conformal intervals, rolling origin, DM tests, drift
  tracking.py                Release provenance and artifact hashing
scripts/                     One runnable step per concern (run_fetch, run_features, run_evaluate,
                             run_intervals, run_error_analysis, run_analysis, run_shap,
                             run_hourly, validate_results)
tests/                       pytest suite (framework, features, models, evaluation, scripts)
results/                     Compact tracked manifests plus locally generated evidence tables
report/                      LaTeX technical report (source, markdown, PDF)
docs/                        Technical design notes

Data split (config.yaml): train through 2023-12-31, validation 2024-01-01 to 2025-06-30, calibration 2025-07-01 to 2025-12-31, and retrospective final evaluation 2026-01-01 to 2026-08-04.

The hourly headline uses only horizon-24 rows: each valid hour is forecast exactly 24 hours earlier. It is not a common-origin day-ahead profile in which one issue time produces all 24 hours.

Results

Release results

• Period: 01 Jan 2026 to 04 Aug 2026; n=216 days.
• Daily forecast: Ensemble. MAE 997.7 MW, MAPE 1.855%, and MASE 0.442. Reference models: Naive (t-1) MAPE 7.025% (n=216); Seasonal naive (t-7) MAPE 5.575% (n=216).
• Hourly forecast: Residual hybrid. Daily-total alignment reduced MAE from 1624.9 MW to 1562.1 MW (-3.86%). The result contains n=5183 hourly values. The daily model was LightGBM.
• Model comparison: Validation MAE was 1463.6 MW for Residual hybrid and 1466.6 MW for Direct LightGBM. The paired difference on the final data was +1.6 MW. Its 95% range was [-44.5, 46.7] MW across 216 days. This range includes zero, so the data do not show a clear winner.
• Uncertainty ranges: Symmetric 90%: target 90%, measured coverage 85.18%, mean width 5612.0 MW, interval score 8837.3 MW, n=5183; Adaptive 90%: target 90%, measured coverage 89.95%, mean width 6545.0 MW, interval score 8273.5 MW, n=5183; CQR 90%: target 90%, measured coverage 86.67%, mean width 7556.4 MW, interval score 10519.8 MW, n=5183.
• Scope: The final period was already examined. These results describe this data period and do not state future accuracy.

Documentation

• docs/TECHNICAL.md — Technical design: system flow, validation rules, and protocol constraints.
• report/technical-report-en.md — Full technical report (also .tex and PDF).
• DATA_LICENSE.md — License notes: code is MIT; data files retain their provider licenses.

License

MIT. See LICENSE. Code is MIT; data files retain their provider licenses (see DATA_LICENSE.md).