feat(ml): initialise le pipeline d'entrainement LightGBM (ADR 0005)

ml/tests/test_train.py

@@ -0,0 +1,76 @@
+from datetime import UTC, datetime, timedelta
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+from enervision_ml.features import TARGET_COLUMN, build_features, feature_columns
+from enervision_ml.train import chronological_split, prepare_dataset, train
+
+
+def make_frame(site_id: str, *, heures: int, depart: datetime) -> pd.DataFrame:
+    instants = [depart + timedelta(hours=h) for h in range(heures)]
+    rng = np.random.default_rng(42)
+
+    return pd.DataFrame(
+        {
+            "site_id": site_id,
+            "timestamp": instants,
+            TARGET_COLUMN: 100.0 + 10.0 * np.sin(np.arange(heures) / 24) + rng.normal(0, 1, heures),
+            "temperature_celsius": 15.0,
+            "humidity_percent": 50.0,
+            "solar_irradiance_wm2": 0.0,
+            "is_working_hours": True,
+            "site_type": "office",
+            "capacity_kw": 100.0,
+        }
+    )
+
+
+def test_chronological_split_puts_the_most_recent_rows_in_validation() -> None:
+    depart = datetime(2026, 1, 1, tzinfo=UTC)
+    features = make_frame("site-a", heures=200, depart=depart)
+
+    entrainement, validation = chronological_split(features, test_fraction=0.2)
+
+    assert entrainement["timestamp"].max() < validation["timestamp"].min()
+    # La coupure vient d'un quantile sur les dates : une approximation du taux demande, pas un
+    # decompte exact de lignes.
+    assert abs(len(validation) - 0.2 * len(features)) <= 2
+
+
+def test_prepare_dataset_types_site_type_as_a_pandas_category() -> None:
+    depart = datetime(2026, 1, 1, tzinfo=UTC)
+    features = build_features(make_frame("site-a", heures=200, depart=depart))
+
+    X, y = prepare_dataset(features, feature_columns())
+
+    assert X["site_type"].dtype.name == "category"
+    assert y.name == TARGET_COLUMN
+
+
+def test_train_runs_end_to_end_on_synthetic_data_and_beats_a_dummy_baseline(
+    tmp_path: Path,
+) -> None:
+    depart = datetime(2026, 1, 1, tzinfo=UTC)
+    frame = pd.concat(
+        [
+            make_frame("site-a", heures=400, depart=depart),
+            make_frame("site-b", heures=400, depart=depart),
+        ],
+        ignore_index=True,
+    )
+    csv_path = tmp_path / "synthetic.csv"
+    frame.to_csv(csv_path, index=False)
+
+    model_metrics, baseline_metrics = train(
+        csv_path=csv_path,
+        model_output=tmp_path / "model.txt",
+        test_fraction=0.2,
+        tracking_uri=f"sqlite:///{tmp_path / 'mlflow.db'}",
+    )
+
+    assert (tmp_path / "model.txt").exists()
+    assert model_metrics["n_observations"] > 0
+    assert model_metrics["mae"] >= 0
+    assert baseline_metrics["n_observations"] == model_metrics["n_observations"]