openclimatefix/nowcasting_utils

"Best" and "worst" for each metric

JackKelly opened this issue · 0 comments

To interpret each metric, we need to understand how the metric reflects the "best possible" performance, and the "worst" performance.

"Worst" could be "smart climatology" (an ML model which only gets time of day, day of year, angle & azimuth of the Sun, PV system ID, clearsky irradiance, and maybe historical PV yield). But gets no information at all about clouds.

The "best" could be a model which cheats and "sees the future". i.e. it gets observations (except for the target PV system) across the entire "forecast" horizon. i.e., if it's "forecasting" 12pm to 1pm, it will get actual PV yield from neighbouring PV systems from 12pm to 1pm, along with actual satellite imagery, NWPs, etc for 12pm to 1pm.