Something else
As part of supporting scores aggregation, in particular Hinton plots, we might want to resample some variables prior to computing the scores metric, both in model and observations. An example of this is temperature_at_screen_level. If the output is hourly, currently, the Hinton plot will produce a triangle for every hour, which for a 48h forecast produces a very stretched and hard to read plot. We might want to resample to every 6 hours, by taking the average. For rainfall, we might want to produce 6h rain accumulations.
Non feasible alternative: aggregating within the Hinton plot operator, but then we have challenges with averaging over scores metrics and significance, which may not be scientifically valid. It is much better to aggregate the raw data first, prior to computing any metrics.
Currently, aggregate.time_aggregate only supports a cube. We should make it support a cubelist, where it iterates over and does the sampling for each cube. We should also support an option where it just returns the same cube if the timedelta is zero (so we can use it as standard across more recipes).
Example implementation on branch 2409_rmse_agg_hinton, https://github.com/MetOffice/CSET/pull/2419/changes#diff-05cfcf0a4d084ae6a866040193ee8f83127a832011b334ea5418ddef8801b134
Something else
As part of supporting scores aggregation, in particular Hinton plots, we might want to resample some variables prior to computing the scores metric, both in model and observations. An example of this is
temperature_at_screen_level. If the output is hourly, currently, the Hinton plot will produce a triangle for every hour, which for a 48h forecast produces a very stretched and hard to read plot. We might want to resample to every 6 hours, by taking the average. For rainfall, we might want to produce 6h rain accumulations.Non feasible alternative: aggregating within the Hinton plot operator, but then we have challenges with averaging over scores metrics and significance, which may not be scientifically valid. It is much better to aggregate the raw data first, prior to computing any metrics.
Currently,
aggregate.time_aggregateonly supports a cube. We should make it support a cubelist, where it iterates over and does the sampling for each cube. We should also support an option where it just returns the same cube if the timedelta is zero (so we can use it as standard across more recipes).Example implementation on branch 2409_rmse_agg_hinton, https://github.com/MetOffice/CSET/pull/2419/changes#diff-05cfcf0a4d084ae6a866040193ee8f83127a832011b334ea5418ddef8801b134