Lagged demand and wind dominate
Top 10 features by total gain across all three quantile boosters. demand_lag_30m and wind_forecast_3h together account for ~58% of split gain — unsurprising given GB's 30 GW of wind capacity.
7SDG 7 · affordable and clean energy
GB carbon intensity forecast
deployed API live
A half-hourly forecast of UK grid carbon intensity you can actually run a dishwasher against. Trained on 4.2M rows of half-hourly grid data; deployed on a K3s cluster behind a tiny FastAPI service.
02 · how it works
From the National Grid to your dishwasher
Five stages, retrained nightly. Each box is a small thing that does one job, and fails loudly when it cannot.
01 · INGEST
Data
NESO API, REST, 5-min cron. Generation mix, demand, interconnector flow.
02 · FEATURES
Engineering
Lags, rolling means, ECMWF wind & temperature forecasts. pandas, DuckDB.
03 · MODEL
Quantile regressor
LightGBM, pinball loss, P10/P50/P90. Retrained nightly. MLflow tracking.
04 · API
Forecast service
FastAPI on Cloud Run. Cached 30s. Returns JSON quantiles by horizon.
05 · SURFACE
You
This widget, an iOS shortcut, Home Assistant. Open API, CC-BY 4.0.
03 · explainability Fig. 1 · Feature importance (gain) Fig. 2 · Calibration Fig. 3 · Residuals over horizon
What is the model actually doing?
Three diagnostics I check when something looks off. Same figures used internally, embedded straight from the training notebook.
Quantiles hold, just
Empirical coverage of the P10–P90 band over the last 90 days, by hour-of-day. Slight under-coverage between 17:00 and 20:00 — peak demand period, where the model is least confident.
Error grows non-linearly
Mean absolute pinball loss vs forecast horizon, last 30 days. Roughly flat to 4 hours, then climbs steeply — this is why the default horizon button is 4 hours.
04 · stack & source
Built with boring, observable parts
No exotic dependencies. Everything here has been in production somewhere for at least five years. That's the point.