feat(elecprice): serve native 15-minute Tibber prices for the quarter-hour grid

Request QUARTER_HOURLY exchange prices from Tibber and store them at their
native resolution instead of pre-averaging to hourly values. EOS resamples the
stored records onto the optimization grid on demand (key_to_array), so keeping
the native step size lets both the hourly (interval=3600) and the 15-minute
(interval=900) optimizer be fed the correct grid automatically.

- GraphQL: priceInfoRange resolution HOURLY -> QUARTER_HOURLY (last 960).
- _hourly_series -> _normalize_series: dedupe by timestamp (mean) + sort, no
  1h aggregation; add _resolution_seconds (median of timestamp diffs, fallback
  3600s).
- Resolution-agnostic ETS extrapolation: seasonal windows and history
  thresholds are scaled by slots_per_hour, needed forecast length and the
  prediction index step are computed in slots. Hourly behaviour is unchanged
  (slots_per_hour=1 -> 168/24 seasonal periods, hourly steps).
- Tests: replace the 1h-averaging test with resolution-preserving + dedup
  tests, assert QUARTER_HOURLY in the query, add a 15-min end-to-end test
  (native storage stays 15min, slot-based seasonal periods = 96, 15-min
  forecast index). Hourly backward-compat tests stay green unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0181Cp11xQ2VesG1HtuPJKMe

Author: chloepriceless

CHANGELOG.md

@@ -20,6 +20,11 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
   default 3600 s interval keeps the previous hourly behaviour. The new sub-hourly PV
   providers (pvnode, Forecast.Solar, Solcast) feed their native resolution straight
   into the quarter-hour grid.
+  - The Tibber electricity price provider now requests native 15-minute exchange prices
+    (`priceInfoRange(resolution: QUARTER_HOURLY)`) and stores them at their native
+    resolution, so both the hourly and the 15-minute optimizer are fed the correct
+    grid. The seasonal price extrapolation is resolution-agnostic and stays identical
+    at the default hourly resolution.
 
 ## 0.3.0 (2026-03-17)
 

src/akkudoktoreos/prediction/elecpricetibber.py

@@ -34,7 +34,7 @@
             total
           }
         }
-        priceInfoRange(resolution: HOURLY, last: 840) {
+        priceInfoRange(resolution: QUARTER_HOURLY, last: 960) {
           nodes {
             startsAt
             total
@@ -245,13 +245,36 @@ def _parse_data(self, response: TibberGraphQLResponse) -> pd.Series:
 
         return series_data.sort_index()
 
-    def _hourly_series(self, series: pd.Series) -> pd.Series:
-        """Normalize Tibber prices to hourly values for EOS optimization."""
+    def _normalize_series(self, series: pd.Series) -> pd.Series:
+        """Normalize Tibber prices while preserving their native resolution.
+
+        The Tibber API delivers either hourly or quarter-hourly prices. EOS resamples
+        the stored records onto the optimization grid on demand (``key_to_array``), so
+        the provider must keep the native step size (e.g. 15 minutes) instead of
+        pre-aggregating to hourly values. Duplicate timestamps are collapsed (mean) and
+        the series is sorted, but the resolution is left untouched.
+        """
         if series.empty:
             return series
         series = series.sort_index()
         series.index = pd.to_datetime([to_datetime(index).isoformat() for index in series.index])
-        return series.resample("1h").mean().dropna()
+        series = series.groupby(level=0).mean().sort_index()
+        return series.dropna()
+
+    def _resolution_seconds(self, series: pd.Series) -> int:
+        """Infer the native slot size in seconds from the series timestamps.
+
+        Uses the median of the timestamp differences so that a single outlier gap does
+        not distort the result. Falls back to hourly (3600 s) when fewer than two
+        timestamps are available.
+        """
+        if len(series) < 2:
+            return 3600
+        deltas = pd.DatetimeIndex(series.index).to_series().diff().dropna()
+        if deltas.empty:
+            return 3600
+        resolution = int(round(deltas.dt.total_seconds().median()))
+        return resolution if resolution > 0 else 3600
 
     def _cap_outliers(self, data: np.ndarray, sigma: int = 2) -> np.ndarray:
         mean = data.mean()
@@ -272,33 +295,48 @@ def _predict_median(self, history: np.ndarray, hours: int) -> np.ndarray:
         clean_history = self._cap_outliers(history)
         return np.full(hours, np.median(clean_history))
 
-    def _predict_missing_prices(self, history: np.ndarray, hours: int) -> np.ndarray:
-        """Forecast missing future prices from the available hourly history."""
+    def _predict_missing_prices(
+        self, history: np.ndarray, slots: int, slots_per_hour: int
+    ) -> np.ndarray:
+        """Forecast missing future prices from the available history.
+
+        Works on the native resolution of the series: ``slots_per_hour`` scales the
+        hour-based seasonal windows into slot counts, so the seasonal periods and
+        history thresholds stay correct at both hourly (``slots_per_hour == 1``) and
+        quarter-hourly (``slots_per_hour == 4``) resolution.
+        """
         numeric_history = np.asarray(history, dtype=float)
         numeric_history = numeric_history[np.isfinite(numeric_history)]
-        history_hours = len(numeric_history)
+        history_slots = len(numeric_history)
 
-        if history_hours > TIBBER_WEEKLY_SEASONAL_HOURS:
+        weekly_seasonal_slots = TIBBER_WEEKLY_SEASONAL_HOURS * slots_per_hour
+        daily_seasonal_slots = TIBBER_DAILY_SEASONAL_HOURS * slots_per_hour
+
+        if history_slots > weekly_seasonal_slots:
             logger.info(
                 "Using weekly seasonal ETS forecast for Tibber electricity prices "
-                "with {} historical hourly values.",
-                history_hours,
+                "with {} historical values.",
+                history_slots,
+            )
+            return self._predict_ets(
+                numeric_history, seasonal_periods=168 * slots_per_hour, hours=slots
             )
-            return self._predict_ets(numeric_history, seasonal_periods=168, hours=hours)
-        if history_hours > TIBBER_DAILY_SEASONAL_HOURS:
+        if history_slots > daily_seasonal_slots:
             logger.info(
                 "Using daily seasonal ETS forecast for Tibber electricity prices "
-                "with {} historical hourly values.",
-                history_hours,
+                "with {} historical values.",
+                history_slots,
             )
-            return self._predict_ets(numeric_history, seasonal_periods=24, hours=hours)
-        if history_hours > 0:
+            return self._predict_ets(
+                numeric_history, seasonal_periods=24 * slots_per_hour, hours=slots
+            )
+        if history_slots > 0:
             logger.warning(
                 "Using median fallback for Tibber electricity prices because only {} "
-                "historical hourly values are available.",
-                history_hours,
+                "historical values are available.",
+                history_slots,
             )
-            return self._predict_median(numeric_history, hours=hours)
+            return self._predict_median(numeric_history, hours=slots)
 
         logger.error("No data available for prediction")
         raise ValueError("No data available")
@@ -310,16 +348,20 @@ def _update_data(self, force_update: Optional[bool] = False) -> None:
             raise ValueError(f"Start DateTime not set: {self.ems_start_datetime}")
 
         api_history_count, api_today_count, api_tomorrow_count = self._api_price_counts(tibber_data)
-        series_data = self._hourly_series(self._parse_data(tibber_data))
+        series_data = self._normalize_series(self._parse_data(tibber_data))
         if series_data.empty:
-            raise ValueError("Tibber response contains no usable hourly price points")
+            raise ValueError("Tibber response contains no usable price points")
+
+        resolution_seconds = self._resolution_seconds(series_data)
+        slots_per_hour = round(3600 / resolution_seconds)
 
         highest_orig_datetime = to_datetime(series_data.index.max())
         self.key_from_series("elecprice_marketprice_wh", series_data)
 
         history = self.key_to_array(
             key="elecprice_marketprice_wh",
             end_datetime=highest_orig_datetime,
+            interval=to_duration(f"{resolution_seconds} seconds"),
             fill_method="linear",
         )
 
@@ -328,36 +370,41 @@ def _update_data(self, force_update: Optional[bool] = False) -> None:
             logger.error(error_msg)
             raise ValueError(error_msg)
 
-        needed_hours = int(
-            self.config.prediction.hours
-            - ((highest_orig_datetime - self.ems_start_datetime).total_seconds() // 3600)
-        )
+        covered_slots = (
+            highest_orig_datetime - self.ems_start_datetime
+        ).total_seconds() // resolution_seconds
+        needed_slots = int(self.config.prediction.hours * slots_per_hour - covered_slots)
 
-        if needed_hours <= 0:
+        if needed_slots <= 0:
             logger.warning(
                 "No prediction needed. "
-                f"needed_hours={needed_hours}, "
+                f"needed_slots={needed_slots}, "
                 f"hours={self.config.prediction.hours}, "
+                f"slots_per_hour={slots_per_hour}, "
                 f"highest_orig_datetime={highest_orig_datetime}, "
                 f"start_datetime={self.ems_start_datetime}"
             )
             return
 
         logger.info(
-            "Tibber electricity price input: api_history_hours={}, api_today_hours={}, "
-            "api_tomorrow_hours={}, combined_history_hours={}, needed_forecast_hours={}.",
+            "Tibber electricity price input: api_history={}, api_today={}, "
+            "api_tomorrow={}, resolution_seconds={}, combined_history_slots={}, "
+            "needed_forecast_slots={}.",
             api_history_count,
             api_today_count,
             api_tomorrow_count,
+            resolution_seconds,
             len(history),
-            needed_hours,
+            needed_slots,
+        )
+        prediction = self._predict_missing_prices(
+            history, slots=needed_slots, slots_per_hour=slots_per_hour
         )
-        prediction = self._predict_missing_prices(history, hours=needed_hours)
 
         prediction_series = pd.Series(
             data=prediction,
             index=[
-                highest_orig_datetime + to_duration(f"{i + 1} hours")
+                highest_orig_datetime + to_duration(f"{(i + 1) * resolution_seconds} seconds")
                 for i in range(len(prediction))
             ],
         )

tests/test_elecpricetibber.py

@@ -175,14 +175,41 @@ def test_parse_data_combines_sorts_and_converts_total(provider, tibber_response)
     assert series.iloc[2] == pytest.approx(0.00030468)
 
 
-def test_tibber_hourly_series_averages_quarter_hour_prices(provider):
-    """Quarter-hour Tibber prices are averaged to hourly EOS prices."""
+def test_tibber_normalize_series_preserves_quarter_hour_resolution(provider):
+    """Quarter-hour Tibber prices keep their native 15-min resolution (no averaging).
+
+    EOS resamples onto the optimization grid on demand, so the provider must store the
+    native step size instead of pre-aggregating quarter-hour prices to hourly values.
+    """
     index = pd.date_range("2026-07-09T00:00:00+00:00", periods=8, freq="15min")
-    series = pd.Series([0.10, 0.30, 0.50, 0.70, 1.0, 1.4, 1.8, 2.2], index=index)
+    values = [0.10, 0.30, 0.50, 0.70, 1.0, 1.4, 1.8, 2.2]
+    series = pd.Series(values, index=index)
+
+    normalized = provider._normalize_series(series)
+
+    # Every 15-min point survives, values untouched, still on a 15-min grid.
+    assert normalized.tolist() == pytest.approx(values)
+    deltas = normalized.index.to_series().diff().dropna().dt.total_seconds().unique().tolist()
+    assert deltas == [900.0]
+    assert provider._resolution_seconds(normalized) == 900
+
+
+def test_tibber_normalize_series_deduplicates_timestamps(provider):
+    """Duplicate timestamps are collapsed (mean) without changing the resolution."""
+    index = pd.DatetimeIndex(
+        [
+            "2026-07-09T00:00:00+00:00",
+            "2026-07-09T00:00:00+00:00",
+            "2026-07-09T01:00:00+00:00",
+        ]
+    )
+    series = pd.Series([0.10, 0.30, 0.50], index=index)
 
-    hourly = provider._hourly_series(series)
+    normalized = provider._normalize_series(series)
 
-    assert hourly.tolist() == pytest.approx([0.40, 1.60])
+    assert len(normalized) == 2
+    assert normalized.iloc[0] == pytest.approx(0.20)
+    assert normalized.iloc[1] == pytest.approx(0.50)
 
 
 def test_empty_tomorrow_stores_only_today_and_warns(provider):
@@ -223,6 +250,7 @@ def test_request_forecast_uses_tibber_graphql_api(
     assert "query" in kwargs["json"]
     assert "TibberPriceInfo" in kwargs["json"]["query"]
     assert "priceInfoRange" in kwargs["json"]["query"]
+    assert "QUARTER_HOURLY" in kwargs["json"]["query"]
     assert "total" in kwargs["json"]["query"]
     assert kwargs["timeout"] == 30
 
@@ -299,3 +327,62 @@ def fake_predict_ets(history, seasonal_periods, hours):
 
     assert forecast_call["seasonal_periods"] == 168
     assert forecast_call["history_hours"] > 840
+
+
+def test_tibber_update_preserves_quarter_hour_resolution_and_slots(
+    tibber_provider, monkeypatch
+):
+    """15-minute Tibber prices are stored natively and extrapolated on the slot grid.
+
+    Proves the resolution-agnostic path: (a) the native 15-min resolution survives
+    storage, (b) the ETS extrapolation scales the seasonal window into slots
+    (daily-only history -> 24*4 = 96 seasonal periods), and (c) the forecast index is
+    spaced at 15-minute steps.
+    """
+    data = TibberGraphQLResponse.model_validate(
+        _tibber_payload(
+            [
+                _price("2026-07-09T00:00:00+00:00", 0.30),
+                _price("2026-07-09T00:15:00+00:00", 0.42),
+                _price("2026-07-09T00:30:00+00:00", 0.36),
+            ],
+            include_history_range=False,
+        )
+    )
+    monkeypatch.setattr(tibber_provider, "_request_forecast", lambda **_: data)
+    forecast_call = {}
+
+    def fake_predict_ets(history, seasonal_periods, hours):
+        forecast_call["seasonal_periods"] = seasonal_periods
+        forecast_call["history_slots"] = len(history)
+        forecast_call["forecast_slots"] = hours
+        return np.full(hours, 0.0009)
+
+    monkeypatch.setattr(tibber_provider, "_predict_ets", fake_predict_ets)
+
+    # A bit more than one week of quarter-hour history: enough for the daily seasonal
+    # window (> 24*7*4 = 672 slots) but below the weekly one (<= 24*35*4 = 3360 slots).
+    stored_history = pd.Series(
+        data=np.linspace(0.0002, 0.0004, 800),
+        index=pd.date_range("2026-07-01T00:00:00+00:00", periods=800, freq="15min"),
+    )
+    tibber_provider.key_from_series("elecprice_marketprice_wh", stored_history)
+
+    tibber_provider._update_data(force_update=True)
+
+    # (b) Daily seasonal window scaled into 15-min slots.
+    assert forecast_call["seasonal_periods"] == 96
+    assert 672 < forecast_call["history_slots"] <= 3360
+    # prediction.hours (6) * slots_per_hour (4) - covered slots (2 -> 00:00..00:30) = 22
+    assert forecast_call["forecast_slots"] == 22
+
+    # (a)+(c) Stored records keep the native 15-min grid across today and the forecast.
+    stored = tibber_provider.key_to_series(
+        "elecprice_marketprice_wh",
+        start_datetime=to_datetime("2026-07-09T00:00:00+00:00"),
+        end_datetime=to_datetime("2026-07-09T06:15:00+00:00"),
+    )
+    steps = stored.index.to_series().diff().dropna().dt.total_seconds().unique().tolist()
+    assert steps == [900.0]
+    # 00:00..06:00 inclusive at 15-min steps = 25 points (3 API + 22 forecast).
+    assert len(stored) == 25