Multiple Datasets III - Configure time coverage

[JPXKQX]

Add the ability to define time coverage individually per dataset/source, both in input and output.

Goal

Train a forecasting model that:

• Takes as input: ERA [t, t-6h] + CERRA [t, t-1h, t-2h]
• Takes as output: cerra [t+1h, t+2h, t+3h]
• No rollout
• No diagnostics

Scope:

• Config: Extend dataset config to allow specifying custom temporal offsets for each dataset.
• Data module: Implement logic in datamodule/data handler to fetch time slices accordingly.
• Training: Adapt the loss function to be computed over all time steps.

Notes

• A step toward native support for temporal interpolation, delayed/coupled inputs, multi-step output, ...
• Requires tight integration with metadata and config schema. New Metadata Schema #229

[eliott-lumet]

Hello, for the spatial downscaling task (#233) it would be relevant to also handle the training of a model that:

• Takes as input: ERA [t]
• Takes as output: CERRA [t]

[mchantry]

Thanks @eliott-lumet that's helpful feedback. Downscaling is planning for this milestone, through a combination of this issue and #233

[sabrinawahl]

For regional models an additional feature would be to have a model that takes global/boundary data at future/predicted time steps as input:

• Global data input at t, t+3h
• Regional data input t, t-3h, ...
• Regional data output t+3h