New north correction method

.gitattributes

@@ -2,3 +2,4 @@ tests/test_data/smarteole/* filter=lfs diff=lfs merge=lfs -text
 tests/test_data/wedowind/Turbine_Upgrade_Dataset.zip filter=lfs diff=lfs merge=lfs -text
 tests/test_data/hot/scada/* filter=lfs diff=lfs merge=lfs -text
 tests/test_data/hot/reanalysis_data/ERA5T_57.50N_-3.25E_100m_1hr_20241231.parquet filter=lfs diff=lfs merge=lfs -text
+tests/test_data/hot/northing/*.parquet filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -86,3 +86,7 @@ output/
 # matplotlib testing
 **/result_images/
 /docs/v1/extra_docs/
+
+# local working documents, not part of the published project
+/CLAUDE.md
+/docs/superpowers/

benchmarking/baselines/example_prepost_study.py

@@ -41,6 +41,7 @@
 from benchmarking.baselines.v0_binned import V0BinnedMethod
 from benchmarking.harness import Method, StudyConfig, leaderboard, plot_campaign_curves, score_study
 from benchmarking.harness.example_hot_study import OracleMethod
+from benchmarking.harness.northing import era5_direction
 from benchmarking.synthetic import HOT_COLUMNS, HOT_RATED_POWER_KW
 from benchmarking.synthetic.make_example_datasets import example_profiles
 from benchmarking.synthetic.sources.hill_of_towie import load_hot_scada
@@ -116,6 +117,11 @@ def run_prepost_study(
     out_dir.mkdir(parents=True, exist_ok=True)
 
     context = build_hot_v0_context(data_dir=data_dir, wtg_names=DEFAULT_TURBINE_SUBSET)
+    # The shared northing step runs per replicate, so every method sees a north-calibrated
+    # direction that wind-up discovered for itself rather than one supplied to it.
+    era5_wd = era5_direction(
+        context.reanalysis_datasets[0].data, pd.DatetimeIndex(base_scada.index.unique()).sort_values()
+    )
     scratch_dir = out_dir / "windup_runs"
 
     all_results = []
@@ -159,6 +165,7 @@ def run_prepost_study(
             methods=methods,
             study=study,
             profile_name=profile_name,
+            era5_wd=era5_wd,
             on_method_complete=partial(save_per_method_curve, out_dir, profile_name),
         )
         summary = leaderboard(results)

benchmarking/baselines/example_toggle_study.py

@@ -42,6 +42,7 @@
 from benchmarking.baselines.v0_binned import V0BinnedMethod
 from benchmarking.harness import Method, StudyConfig, leaderboard, plot_campaign_curves, score_study
 from benchmarking.harness.example_hot_study import OracleMethod
+from benchmarking.harness.northing import era5_direction
 from benchmarking.synthetic import HOT_COLUMNS, HOT_RATED_POWER_KW, ConstantCpChange
 from benchmarking.synthetic.sources.hill_of_towie import load_hot_scada
 
@@ -89,6 +90,11 @@ def run_toggle_study(
     out_dir.mkdir(parents=True, exist_ok=True)
 
     context = build_hot_v0_context(data_dir=data_dir, wtg_names=DEFAULT_TURBINE_SUBSET)
+    # The shared northing step runs per replicate, so every method sees a north-calibrated
+    # direction that wind-up discovered for itself rather than one supplied to it.
+    era5_wd = era5_direction(
+        context.reanalysis_datasets[0].data, pd.DatetimeIndex(base_scada.index.unique()).sort_values()
+    )
     scratch_dir = out_dir / "windup_runs"
 
     all_results = []
@@ -127,6 +133,7 @@ def run_toggle_study(
             methods=methods,
             study=study,
             profile_name=profile_name,
+            era5_wd=era5_wd,
             on_method_complete=partial(save_per_method_curve, out_dir, profile_name),
         )
         summary = leaderboard(results)

benchmarking/baselines/inspect_prepost_hard_case.py

@@ -40,7 +40,7 @@
     MIN_PRE_MONTHS,
     default_output_root,
 )
-from benchmarking.baselines.hot_context import build_hot_v0_context
+from benchmarking.baselines.hot_context import HotV0Context, build_hot_v0_context
 from benchmarking.baselines.naive_ratio import NaiveRatioMethod
 from benchmarking.baselines.overnight_common import start_overnight_run
 from benchmarking.baselines.overnight_profiles import overnight_profiles
@@ -61,6 +61,7 @@
     treated_activity_mask,
     window_row_mask,
 )
+from benchmarking.harness.northing import era5_direction
 from benchmarking.synthetic import HOT_COLUMNS, HOT_RATED_POWER_KW
 from benchmarking.synthetic.sources.hill_of_towie import load_hot_scada
 
@@ -110,11 +111,17 @@ def _select_replicate(replicates: list[Replicate], test_wtg: str) -> Replicate:
 
 
 def _pin_case(
-    scada_df: pd.DataFrame, *, study: StudyConfig, profile_name: str, test_wtg: str, campaign_months: int
+    scada_df: pd.DataFrame,
+    *,
+    study: StudyConfig,
+    profile_name: str,
+    test_wtg: str,
+    campaign_months: int,
+    era5_wd: pd.Series,
 ) -> tuple[Replicate, MethodInput, float, CampaignWindow]:
     """Build the pinned replicate, its shared ``MethodInput``, the ground-truth uplift, and the window."""
     profile = overnight_profiles()[profile_name]
-    replicates = build_replicates(scada_df, profile=profile, study=study)
+    replicates = build_replicates(scada_df, profile=profile, study=study, era5_wd=era5_wd)
     rep = _select_replicate(replicates, test_wtg)
 
     windows = campaign_windows(
@@ -169,13 +176,12 @@ def _power_model(out_dir: Path, era5_hourly_df: pd.DataFrame, *, save_plots: boo
     )
 
 
-def _build_methods(out_dir: Path, *, include_v0: bool) -> list[Method]:
+def _build_methods(out_dir: Path, *, context: HotV0Context, include_v0: bool) -> list[Method]:
     """Return the methods to inspect, each writing diagnostics (plots on) into its own subfolder.
 
     One ``power_model`` run folder: with conditional uplift on (the default) it carries the overall
     diagnostics plus the conditional CSVs (``conditional/``) and the step-7 implied-shrinkage plot.
     """
-    context = build_hot_v0_context(wtg_names=DEFAULT_TURBINE_SUBSET)
     era5 = context.reanalysis_datasets[0].data
     methods: list[Method] = [
         NaiveRatioMethod(
@@ -293,10 +299,18 @@ def inspect_prepost_hard_case(
         wtg_names=DEFAULT_TURBINE_SUBSET,
     )
 
+    context = build_hot_v0_context(wtg_names=DEFAULT_TURBINE_SUBSET)
     rep, mi, truth, window = _pin_case(
-        scada_df, study=study, profile_name=profile_name, test_wtg=test_wtg, campaign_months=campaign_months
+        scada_df,
+        study=study,
+        profile_name=profile_name,
+        test_wtg=test_wtg,
+        campaign_months=campaign_months,
+        era5_wd=era5_direction(
+            context.reanalysis_datasets[0].data, pd.DatetimeIndex(scada_df.index.unique()).sort_values()
+        ),
     )
-    methods = _build_methods(out_dir, include_v0=include_v0)
+    methods = _build_methods(out_dir, context=context, include_v0=include_v0)
     summary, outputs = _run_methods(methods, mi=mi, truth=truth)
 
     _plot_conditional_uplift(

benchmarking/baselines/inspect_wake_steering_case.py

@@ -69,6 +69,7 @@
     condition_bins,
     plot_conditional_uplift,
 )
+from benchmarking.harness.northing import era5_direction, north_scada
 from benchmarking.synthetic import (
     HOT_COLUMNS,
     HOT_RATED_POWER_KW,
@@ -438,6 +439,18 @@ def inspect_wake_steering_case(
         era5 = context.reanalysis_datasets[0].data
 
         full_dataset, steering = _build_dataset(scada_df, metadata_df)
+        # The shared northing step, discovering as it does on every other path -- the methods read
+        # a north-calibrated direction wind-up worked out, not the table the injection was gated on.
+        full_dataset = replace(
+            full_dataset,
+            synthetic_df=north_scada(
+                full_dataset.synthetic_df,
+                columns=HOT_COLUMNS,
+                north_offsets=None,
+                rated_power_kw=HOT_RATED_POWER_KW,
+                era5_wd=era5_direction(era5, pd.DatetimeIndex(full_dataset.synthetic_df.index.unique()).sort_values()),
+            ),
+        )
         schedule = ToggleSchedule(period=TOGGLE_PERIOD, start=TOGGLE_START)
 
         nadir, half_width = _pair_sector(steering, margin_deg=wd_margin_deg)