openstef-models 4.4.2__tar.gz → 4.4.4__tar.gz

{openstef_models-4.4.2 → openstef_models-4.4.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: openstef-models
-Version: 4.4.2
+Version: 4.4.4
 Summary: Core models for OpenSTEF
 Project-URL: Documentation, https://openstef.github.io/openstef/index.html
 Project-URL: Homepage, https://lfenergy.org/projects/openstef/

{openstef_models-4.4.2 → openstef_models-4.4.4}/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-models"
-version = "4.4.2"
+version = "4.4.4"
 description = "Core models for OpenSTEF"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]

{openstef_models-4.4.2 → openstef_models-4.4.4}/src/openstef_models/models/forecasting/flatliner_forecaster.py

@@ -8,7 +8,8 @@ Provides basic forecasting model that predict constant flatliner zero values. It
 when a flatline (non-)zero measurement is observed in the past and expected in the future.
 """
 
-from typing import override
+from datetime import timedelta
+from typing import cast, override
 
 import pandas as pd
 from pydantic import Field, PrivateAttr
@@ -24,8 +25,9 @@ MODEL_CODE_VERSION = 1
 class FlatlinerForecaster(Forecaster, ExplainableForecaster, ContributionsMixin):
     """Flatliner forecaster that predicts a flatline of zeros or median.
 
-    A simple forecasting model that always predicts zero (or the median of historical
-    load measurements if configured) for all horizons and quantiles.
+    A simple forecasting model that always predicts zero (or a median load value if
+    configured) for all horizons and quantiles. The median can be calculated over
+    a trailing window when the fallback is responding to a recent flatline.
 
     Invariants:
         - Configuration quantiles determine the number of prediction outputs
@@ -49,6 +51,10 @@ class FlatlinerForecaster(Forecaster, ExplainableForecaster, ContributionsMixin)
         default=False,
         description="If True, predict the median of load measurements instead of zero.",
     )
+    median_window: timedelta | None = Field(
+        default=None,
+        description="Optional trailing window for the median load estimate; None uses all available measurements.",
+    )
 
     hyperparams: HyperParams = Field(
         default_factory=HyperParams,
@@ -76,7 +82,13 @@ class FlatlinerForecaster(Forecaster, ExplainableForecaster, ContributionsMixin)
         data_val: ForecastInputDataset | None = None,
     ) -> None:
         if self.predict_median:
-            self._median_value = float(data.target_series.median())
+            target_series = data.target_series
+            if self.median_window is not None:
+                last_valid_index = cast(pd.Timestamp | None, target_series.last_valid_index())
+                if last_valid_index is not None:
+                    window_start = last_valid_index - self.median_window
+                    target_series = target_series.loc[window_start:last_valid_index]
+            self._median_value = float(target_series.median())
 
     @override
     def predict(self, data: ForecastInputDataset) -> ForecastDataset:

{openstef_models-4.4.2 → openstef_models-4.4.4}/src/openstef_models/presets/forecasting_workflow.py

@@ -566,6 +566,7 @@ def create_forecasting_workflow(
             quantiles=config.quantiles,
             horizons=config.horizons,
             predict_median=config.predict_nonzero_flatliner,
+            median_window=config.flatliner_threshold,
         )
         postprocessing = [
             QuantileSorter(),

{openstef_models-4.4.2 → openstef_models-4.4.4}/tests/unit/models/forecasting/test_flatliner_forecaster.py

@@ -10,6 +10,8 @@ import pytest
 from openstef_core.datasets.validated_datasets import ForecastInputDataset
 from openstef_core.types import LeadTime, Quantile
 from openstef_models.models.forecasting.flatliner_forecaster import FlatlinerForecaster
+from openstef_models.models.forecasting_model import ForecastingModel
+from openstef_models.presets import ForecastingWorkflowConfig, create_forecasting_workflow
 
 
 @pytest.fixture
@@ -64,3 +66,55 @@ def test_predict_returns_median_when_predict_median_is_true(sample_forecast_inpu
     assert isinstance(result.data, pd.DataFrame)
     assert (result.data == expected_median).all().all()
     assert set(result.data.columns) == {q.format() for q in forecaster.quantiles}
+
+
+@pytest.mark.parametrize(
+    ("load_values", "median_window", "expected_median"),
+    [
+        pytest.param([100.0, 110.0, 120.0, 7.0, 7.0, 7.0], timedelta(hours=2), 7.0, id="repeated-value"),
+        pytest.param([100.0, 110.0, 120.0, 6.0, 8.0], timedelta(hours=1), 7.0, id="two-different-values"),
+    ],
+)
+def test_predict_returns_recent_median_when_median_window_is_set(
+    load_values: list[float],
+    median_window: timedelta,
+    expected_median: float,
+) -> None:
+    # Arrange
+    index = pd.date_range("2025-01-01", periods=len(load_values), freq="h")
+    data = ForecastInputDataset(
+        data=pd.DataFrame({"load": load_values}, index=index),
+        sample_interval=timedelta(hours=1),
+    )
+    forecaster = FlatlinerForecaster(
+        quantiles=[Quantile(0.5)],
+        horizons=[LeadTime(timedelta(hours=1))],
+        predict_median=True,
+        median_window=median_window,
+    )
+
+    # Act
+    forecaster.fit(data)
+    result = forecaster.predict(data)
+
+    # Assert
+    assert (result.data == expected_median).all().all()
+
+
+def test_flatliner_workflow_uses_threshold_for_median_window() -> None:
+    # Arrange
+    config = ForecastingWorkflowConfig(
+        model_id="flatliner-test",
+        model="flatliner",
+        flatliner_threshold=timedelta(hours=6),
+        predict_nonzero_flatliner=True,
+        mlflow_storage=None,
+    )
+
+    # Act
+    workflow = create_forecasting_workflow(config)
+
+    # Assert
+    assert isinstance(workflow.model, ForecastingModel)
+    assert isinstance(workflow.model.forecaster, FlatlinerForecaster)
+    assert workflow.model.forecaster.median_window == timedelta(hours=6)