One model per horizon (1 to 24h), trained once on the data available before
2023-11-25 (the training set of the cross-validation fold that contains the
first date; it ends 7 days before that fold's test period). The same fitted models are then
used to predict 10 dates of that fold's test period, so all the forecasts below are out of
sample. The recalibrated model is calibration.Recalibrator (last 4
inner folds), applied per horizon.
For each date: the grey line is the observed load before the prediction time, the black line the observed load after it, and the coloured band the 0.05–0.95 interval. Drag a rectangle to zoom (both axes), double-click to reset.
| prediction date (UTC) | raw interval contains (of 24 hours) |
recalibrated interval contains | outside raw, inside recalibrated |
|---|---|---|---|
| 2023-12-02 01:00 | 18 | 24 | 6 |
| 2023-12-12 09:00 | 22 | 24 | 2 |
| 2023-12-22 17:00 | 10 | 23 | 13 |
| 2024-01-02 02:00 | 11 | 12 | 1 |
| 2024-01-12 10:00 | 23 | 24 | 1 |
| 2024-01-21 18:00 | 16 | 23 | 7 |
| 2024-02-01 02:00 | 10 | 13 | 3 |
| 2024-02-11 11:00 | 18 | 22 | 4 |
| 2024-02-21 19:00 | 11 | 15 | 4 |
| 2024-03-02 03:00 | 15 | 19 | 4 |