Fail clearly on insufficient forecasting history by translating Darts fit errors

[Copilot]

Forecasting could fail with a low-level Darts ValueError when requested horizon exceeded usable training history.

• Error translation at model-fit boundary

  • Wrap the specific Darts fit-time ValueError (no common times for which features can be created) as NotEnoughDataException.
  • Surface a domain-level message that clearly points to insufficient history vs requested horizon.

• Regression coverage (integration-style)

  • Added an integration-style regression test through TrainPredictPipeline.compute(...) that reproduces the Darts fit failure path and asserts it is converted to NotEnoughDataException.

# train.py (fit boundary)
try:
    model.fit(...)
except ValueError as e:
    if "do not share any common times for which features can be created" in str(e):
        raise NotEnoughDataException(
            "Not enough training data for the requested forecast horizon. "
            "Upload more historical data or choose a shorter forecast duration."
        ) from e
    raise

[BelhsanHmida]

@copilot I don’t think the max-forecast-horizon / predict_period defaulting change belongs in this PR.

Issue #2191 is about translating the Darts “no common feature times” failure into a clear NotEnoughDataException when there is not enough training history. It does not require changing how max-forecast-horizon resolves the prediction window.

Please remove the schema/test changes around preserving the default 48h duration when only max-forecast-horizon is supplied, unless we intentionally want to make that as a separate API/CLI behavior change with its own issue, migration note, and tests.

For this PR, I think the scope should stay focused on:

• catching or pre-detecting the insufficient-history Darts failure
• raising NotEnoughDataException
• adding an integration-style regression test through TrainPredictPipeline.compute(...)

[Copilot]

@copilot I don’t think the max-forecast-horizon / predict_period defaulting change belongs in this PR.

Issue #2191 is about translating the Darts “no common feature times” failure into a...

Addressed in 9183269.

I removed the schema/defaulting behavior change and reverted the related schema test expectations, so this PR no longer changes max-forecast-horizon/predict-period semantics. I also replaced the unit-level wrapping test with an integration-style regression test that goes through TrainPredictPipeline.compute(...) and verifies the Darts ValueError is converted to NotEnoughDataException.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• publicsuffix.org
  • Triggering command: /home/REDACTED/work/flexmeasures/flexmeasures/.venv/bin/pytest pytest exec --process /tmp/runc-process775792770 --detach --pid-file /run/containerd/io.containerd.runtime.v2.task/moby/7e7b651f7c7b1f05c1ef82382934c0adfd6fef18d5796d04855f29c5fd9aaecd/ec319217d5b6a5c16c8e6f1ac2ff80c9ca511327d6a8bf20313263f8c5a10228.pid 7e7b651f7c7b1f05c1ef82382934c0adfd6fef18d5796d04855f29c5fd9aaecd (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

[BelhsanHmida]

@copilot 1. The new test goes through TrainPredictPipeline.compute(...), which is good, but it still monkeypatches CustomLGBM.fit to raise the error. So it proves the wrapping path, but not the real #2191 short-history / long-horizon failure. Can we make this reproduce the actual Darts failure without monkeypatching fit?

[Copilot]

@copilot 1. The new test goes through TrainPredictPipeline.compute(...), which is good, but it still monkeypatches CustomLGBM.fit to raise the error. So it proves the wrapping path, but not the re...

Implemented in 18f55c8.

The regression test no longer monkeypatches CustomLGBM.fit. It now creates a real 15-minute sensor with only 3 days of history and requests a 4-day (PT96H) horizon through TrainPredictPipeline.compute(...), which reproduces the real short-history/long-horizon Darts failure path and asserts it is surfaced as NotEnoughDataException.

[BelhsanHmida]

approved!

[nhoening]

Now what? There should be a second human in the loop right?

[BelhsanHmida]

Now what? There should be a second human in the loop right?

Yes i was thinking to ask second review from yourself or felix.