feat(application): integrate measured-load ingestion training and planner source

2026-10-02 21:00:05 +00:00
parent 44362e4bf5
commit fc024fcfba
96 changed files with 8933 additions and 0 deletions
@@ -0,0 +1,106 @@
+"""Measured telemetry integrity and repeat-profile helpers.
+No IO, no fabricated measurements and no fixed household-load fallback.
+Forecast freshness is checked against raw telemetry, never against filled features.
+"""
+from __future__ import annotations
+import datetime
+import math
+import numpy as np
+import pandas as pd
+
+MEASURED_COLUMNS = ('PV', 'Hausverbrauch', 'Netzleistung', 'SOC')
+
+class TelemetryUnavailable(ValueError):
+    """No publishable forecast can be derived from the supplied observations."""
+
+
+def sanitize_measured_frame(frame):
+    out = frame.copy()
+    for column in MEASURED_COLUMNS:
+        if column not in out:
+            continue
+        raw = out[column]
+        numeric = pd.to_numeric(raw, errors='coerce').astype(float)
+        bad = ~np.isfinite(numeric) | raw.map(lambda x: isinstance(x, (bool, np.bool_)))
+        if column in ('PV', 'Hausverbrauch', 'SOC'):
+            bad |= numeric < 0
+        if column == 'SOC':
+            bad |= numeric > 100
+        out[column] = numeric.mask(bad)
+    return out
+
+
+def _naive_utc(value):
+    stamp = pd.Timestamp(value)
+    if stamp.tzinfo is not None:
+        stamp = stamp.tz_convert('UTC').tz_localize(None)
+    return stamp
+
+
+def require_recent_telemetry(data_obj, fields=('PV','Hausverbrauch'), max_age_minutes=30.0):
+    """Refuse missing/stale inputs without changing real zeros or raw samples."""
+    if isinstance(max_age_minutes, bool) or not math.isfinite(max_age_minutes) or max_age_minutes <= 0:
+        raise ValueError('Positive telemetry age limit required')
+    now = _naive_utc(data_obj['now'])
+    raw = data_obj.get('df_recent_raw')
+    raw = sanitize_measured_frame(raw) if raw is not None else pd.DataFrame()
+    if not raw.empty:
+        raw.index = pd.DatetimeIndex([_naive_utc(t) for t in raw.index])
+        raw = raw.loc[raw.index < now].sort_index()
+    report = {}
+    errors = []
+    for field in fields:
+        if field not in MEASURED_COLUMNS:
+            raise ValueError('Unknown telemetry target')
+        values = raw[field].dropna() if field in raw else pd.Series(dtype=float)
+        if values.empty:
+            errors.append(field + ': keine gemessenen Werte')
+            continue
+        stamp = values.index[-1]
+        age = (now-stamp).total_seconds()/60.0
+        report[field] = {'lastObservedInterval': stamp.isoformat()+'Z', 'ageMinutes': age,
+                         'observedIntervals': int(len(values)), 'lastValue': float(values.iloc[-1])}
+        if age > max_age_minutes:
+            errors.append(field + ': Messdaten veraltet (' + format(age,'.1f') + ' min)')
+    if errors:
+        raise TelemetryUnavailable('; '.join(errors) + '. Keine neuen Prognosen/Fahrplaene veroeffentlicht.')
+    return report
+
+
+def _finite_nonnegative(value):
+    if isinstance(value, (bool, np.bool_)):
+        return None
+    try:
+        value = float(value)
+    except (ValueError, TypeError):
+        return None
+    return value if math.isfinite(value) and value >= 0.0 else None
+
+
+def profile_source_value(history, at, column, predictions=None):
+    """Repeat yesterday; prefer same weekday if yesterday is missing, then older days.
+    Explicitly generated first-day values may be repeated on the second forecast day.
+    Missing historical values are not zeros. No future measurement is read.
+    """
+    if column not in ('PV','Hausverbrauch'):
+        raise ValueError('Unsupported repeat-profile target')
+    predictions = {} if predictions is None else predictions
+    at = pd.Timestamp(at)
+    for day in (1, 7, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14):
+        source = at - datetime.timedelta(days=day)
+        if source in predictions:
+            value = _finite_nonnegative(predictions[source])
+            if value is not None:
+                return value
+        if history is not None and column in history and source in history.index:
+            value = _finite_nonnegative(history.at[source,column])
+            if value is not None:
+                return value
+    raise TelemetryUnavailable(column + ': kein gemessener Tagesprofilwert fuer ' + str(at))
+
+
+def repeat_daily_profile(history, future_index, column):
+    result = {}
+    for at in future_index:
+        result[at] = profile_source_value(history, at, column, result)
+    return result