implement multi-farm support in pywake_api

[kilojoules]

No description provided.

[bjarketol]

Hi @kilojoules,

I did a quick agent-based review. It found a few things.

High

• Per-farm AEP normalization mismatch — wifa/pywake_api.py:1040-1043, 1252-1254. For time-series cases, aep_per_turbine is computed with normalize_probabilities=True while total AEP uses normalize_probabilities=False. Multi-farm per-farm AEP is derived from aep_per_turbine, so it is on a different basis than total AEP. Fix: use the same normalization (False) for per-turbine AEP in time-series, then slice/sum per farm.

• Turbine deduplication by name only — wifa/pywake_api.py:285-293 (_build_multifarm_turbines). Two farms whose turbine YAMLs share a name (e.g. IEA-15MW) but differ in power curve/geometry are silently merged into the first farm's spec — wrong physics, no warning. Fix: dedupe by structural fingerprint, or raise when distinct specs share a name. At minimum warn on reuse.

Medium

• hub_heights collision across farms — wifa/pywake_api.py:97-100, 294, 1234. Single-type farms always key hub_heights with local "0"; hub_heights.update(hh) overwrites prior farms. Distinct hub heights are lost, breaking height-dependent site interpolation / FUGA setup. Fix: key by global turbine-type index (or farm-scoped key) and derive first_hh from merged global turbine objects.

• generate_outputs still receives a list in multi-farm path — wifa/pywake_api.py:~1249. generate_outputs(sim_results, system_dat, ...) was written for a single-farm dict; in the multi-farm branch system_dat["wind_farm"] is a list. Either crashes iterating farm_dat["layouts"] or silently writes only farm 0's artifacts. The new test only checks the return value. Fix: teach generate_outputs to handle a farm list (per-farm outputs), or branch/skip and document the gap. Extend the test to assert output-dir contents.

• farm_dat = farms[0] as a single-farm stand-in — wifa/pywake_api.py:~1218. Inline comment admits "assumes shared turbine spec across farms." Any downstream code reading farm_dat (e.g. rd lookup, layout-derived state) silently uses farm 0 only — defeating the point of multi-farm. Fix: refactor downstream usages to operate per-farm, or assert single-turbine-type in the multi-farm path. Don't paper over with a comment.

• Flaky wall-clock assertion — tests/test_pywake.py:~516 (test_pywake_multifarm_runs_fast). assert elapsed < 10.0 is flaky on shared CI and doesn't validate correctness. Fix: drop the timing assertion (or move to an opt-in benchmark) and rename the test to reflect what it verifies (per-farm AEP + wake interaction).

Low

• Unused _type_names — wifa/pywake_api.py:~284. Unpacked from _build_farm_turbine_list but never used; the type-name → global-index map is rebuilt from wt.name(), inconsistent with the per-farm "0"-style keys. Fix: remove the unused return, or use _type_names consistently.

• layouts shape assumption — wifa/pywake_api.py:~136 (_build_farm_turbine_list). The multi-type branch indexes farm_dat["layouts"][0]["turbine_types"] while the single-type branch handles both list and dict forms; a multi-type config with dict-form layouts crashes with an opaque error. Fix: normalize layouts to a list once at the top of the helper.

[SchmJo]

is this only for pywake? can we setup a new example case, such that I can also adopt foxes to read multiple farms from windio? Thanks

[kilojoules]

Pushed 6ec490f addressing the review findings and adding the requested example case:

• New examples/cases/multiple_wind_farms: the IEA Wind 2200-22-MW Reference Offshore Wind Plant (three farms sharing one site and energy resource, from IEAWindSystems/IEA-Wind-2200-22-ROWP, Apache-2.0). run_pywake returns one AEP per farm, and tests/test_pywake.py::test_pywake_multifarm_rowp_example runs it end to end. The example is tool-agnostic windIO, so foxes can adopt the same input — the runner support here is pywake-only so far; foxes just needs its reader to accept a list of farms.
• Review fixes: turbines are deduplicated across farms by full spec (same name + different spec now raises), hub heights are keyed by global type index instead of colliding per-farm "0" keys, the FUGA rotor diameter comes from the merged specs instead of farms[0], per-turbine AEP uses the same normalization basis as the total (so per-farm AEPs sum to it), and the wall-clock assertion is dropped. generate_outputs never touches wind_farm, so the multi-farm path writes the combined outputs fine; splitting artifacts per farm would be a follow-up.
• Multi-farm wind_farm arrays only validate against the upstream IEAWindSystems/windIO schema (the EUFlow fork predates it), so the dependency pin moved to upstream, and AWAKEN's cutin_wind_speed/cutout_wind_speed moved under performance to conform to that schema.

[kilojoules]

@bjarketol I think my coding agent adressed these things, detailed in the above message. Does this look OK to commit?

[kilojoules]

@SchmJo I added a multi wind farm example

[SchmJo]

@kilojoules Supported by foxes now. Added to foxes tests, and they are green.

[bjarketol]

Second-round agent review (4 agents: fix verification, fresh bug hunt, data/schema/cross-tool, test adequacy), at 952e8cd.

Round 1: all 8 findings verified fixed in 6ec490f — including numerically confirming per-farm AEPs sum to the total on both Weibull and time-series paths, and that generate_outputs never touches wind_farm.

New critical findings

1. Single-farm regression (crash): a farm with a turbine_types dict but no per-position turbine_types in the layout (schema-valid, ran fine on main) now raises KeyError at wifa/pywake_api.py:107. Needs a fallback to type 0 when there's one spec and the layout key is absent.
2. Silent per-farm AEP misattribution: farm_slices are sized by len(layout["turbine_types"]) but x/y concatenation by the coordinate arrays (pywake_api.py:207-209 vs 1131-1134). Probe with 3 coordinates / 2 type entries ran without error — one farm's AEP absorbed a neighbor's turbine, another turbine vanished from all per-farm totals. Needs an explicit length-mismatch ValueError.
3. Type indices resolved by dict position, not key: layout integers are windIO keys into the turbine_types mapping, but the code does list(dict.values()) + positional lookup (pywake_api.py:106-107, 207). 1-based keys (see examples/cases/windio_4turbines_multipleTurbines) → IndexError; unordered keys → silently wrong turbine model. Pre-existing flaw, but this PR rewrote the mapping layer with no coverage.
4. test_foxes_multiple_farms fails in the default env: plain uv run pytest resolves foxes 1.7.2 from PyPI (via wayve), which can't read a wind_farm list — the PR's own test fails with TypeError. Passes only with --extra foxes (1.8.4 from eu_flow). Needs a capability skip or explicit CI extra.
5. wayve and floris crash with a raw TypeError on the multi-farm example (verified). Both should detect the list and raise a clear NotImplementedError; same audit for cs_api.

Medium

• Both new git deps are floating (windIO head-of-main, foxes@eu_flow branch) — only uv.lock protects CI. Pin a commit/tag.
• Time-series multi-hub-height paths crash (now reachable via multi-farm): missing TI leaves site = None (pywake_api.py:467-498); two types at the same hub height hit a dead dedup guard → xarray dim conflict (:471-473).
• docs/source/getting_started/installation.rst:60-80 still instructs installing the EUFLOW windIO fork, whose schema may reject the new example.
• foxes merges all farms into one 100-turbine result with no farm coordinate — per-farm AEP unrecoverable, inconsistent with pywake's return. Document at minimum.
• Weak tests: ROWP test only asserts len(aep) == 3 and positivity; nothing asserts sum(per_farm) == total; multi-farm output artifacts unchecked; test_foxes_multiple_farms asserts nothing. Untested: multi-farm + time-series, dict-form layouts, multi-type single farm.

Low

• run_pywake docstring still says returns float; multi-farm returns list[float].
• Spec-conflict check uses dict !=; numpy arrays from dict-input callers → "ambiguous truth value" instead of the intended ValueError.
• ROWP example dir needs a short README with Apache-2.0 attribution (WIFA is MIT); €/m vs USD/m inconsistency in wind_farm.yaml.

Verified good

Type indexing matches from_WindTurbine_lst order, farm slices align with concatenation order (incl. mixed-type farms), zero-turbine farms work; ROWP example internally consistent; AWAKEN cutin/cutout move is schema-correct and fixes a silent bug (values were previously ignored — AWAKEN AEP numbers will shift).

Recommendation: hold merge until criticals 1–4 are addressed. Highest-value tests: multi-type single farm with 1-based keys, sum(per_farm) == total for both resource types, length-mismatch ValueError.