Disaggregation of temporally coarse forecasts results in NaNs

[jonasbhend]

When disaggregating forecasts with 6hourly resolution to hourly, the necessary input data is not available. This PR fixes the default behaviour (error) to allow plotting of meteograms and dashboards for diverse datasets of forecasters (6-hourly) and temporal downscalers (1-hourly).

Summary of changes

• ensure only available lead times are loaded from grib files
• reindex before cumsum to avoid error

[dnerini]

hi @jonasbhend, the code changes look good to me, but I'm not entirely sure I understand the problem. Is this issue affecting baselines or ML runs? can you provide a concrete example scenario that would trigger it?

[dnerini]

should there be some way to distinguish "expected coarser resolution" from "unexpected missing data"? should this be LOG.warning instead of debug to avoid that real data problems go unnoticed?

[jonasbhend]

hi @jonasbhend, the code changes look good to me, but I'm not entirely sure I understand the problem. Is this issue affecting baselines or ML runs? can you provide a concrete example scenario that would trigger it?

The concrete example is this config with a forecaster and a temporal downscaler being evaluated side-by-side (something we have done before). With explicit aggregations (i.e. TOT_PREC1 and TOT_PREC6), this will result in an error for both the experiment and meteogram use cases:

# yaml-language-server: $schema=../workflow/tools/config.schema.json
description: |
  Evaluate skill of Varda-single-1.0 against ground observations.

config_label: varda-single-1.0

dates:
  start: 2025-03-01T00:00
  end: 2025-03-03T00:00
  frequency: 24h

runs:
  - temporal_downscaler:
      checkpoint: https://service.meteoswiss.ch/mlstore#/models/sruc-m-2-interpolator/versions/3
      label: Varda-single-1.0
      steps: 0/120/1
      config: resources/inference/configs/sgm-temporal-downscaler-global_trimedge_multi.yaml
      extra_requirements:
        - anemoi-datasets==0.5.35
        # NOTE this pins the merge commit that introduced 'copy-prognostic-from-forecaster'
        # TODO pin a release tag once the plugins are versioned. Do not pin the main branch to avoid stale code (changes do to main to not change the evalml hash)
        - git+https://github.com/MeteoSwiss/anemoi-plugins-meteoswiss.git@b4835d76346e1e5ec0181884520da00e317fbfbc
        # - anemoi-inference==0.11.0
      forecaster:
        checkpoint: https://service.meteoswiss.ch/mlstore#/models/sruc-m-1-forecaster/versions/4
        config: resources/inference/configs/sgm-multidataset-forecaster-global-ich1-oper.yaml
        steps: 0/120/6
  - forecaster:
      label: Forecaster
      checkpoint: https://service.meteoswiss.ch/mlstore#/models/sruc-m-1-forecaster/versions/4
      config: resources/inference/configs/sgm-multidataset-forecaster-global-ich1-oper.yaml
      steps: 0/120/6

experiment:
  params:
    - TOT_PREC1
    - TOT_PREC6
...