Undocumented features / forecast accuracy

[doig007]

Undocumented features

Some undocumented features of the API that those using it may wish to consider based upon reviewing what the API has periodically returned:

1. intensity.actual field is populated 30 minutes after the end of each period, but it is liable for revisions at least 24 hours after the end of that period, presumably reflecting Elexon data flows. For example, the figure for the period starting '2021-02-02 23:00:00' was still being revised up to at least 23:15:00 the next day.

2. intensity.forecast field is liable for revisions after the end of the period, although not necessarily in a way that reduces forecast error. For example, more than 5 minutes into the period starting '2021-03-01 09:00:00', the reported forecast was 286 g/kWh; whereas c.5 minutes at the end of the period (09:35:02) the reported forecast was 282 g/kWh. The forecast figure was finally revised sometime between 10:35 and 11:05 (i.e., more than an hour after end of period) to 290 g/kWh, having changed at least once in the meantime.

Forecast accuracy

For those interested in forecast accuracy, based on 2,450 recent periods in Feb-March 2021 in a collected sample (868k individual forecasts), at the beginning of the forecast horizon (4-days ahead) the mean absolute forecast error was c.20g/kWh (equiv. c.11% relative error), which falls gradually until 4 hours before the start of the period, when it drops to c.10g/kWh by start of period. The mean absolute forecast error (and 50th and 90th percentiles) for each lead time up to the period is shown in the figure below.

At day-ahead (-24 hours), the forecast model performs c.46% better than the trivial ‘lagged’ model of assuming day-ahead carbon intensity is identical to current period. However, the volatility structure is such that the forecast model underperforms the lagged model within 2 hours of the period start. This suggests that the forecast model is either unaware of the very recent system condition or it is not giving it adequate weight.

Forcast lead time (hours before period start )	Forecast model mean absolute error (g/kWh)	Lagged model mean absolute error (g/kWh)
72	22.5	62.0
48	19.7	51.4
24	17.4	37.5
2	14.3	16.1
1	12.7	9.0
0.5	11.7	5.1