update for 4 hour rain forecast

scripts/rain_model_common.py

@@ -86,7 +86,46 @@ FEATURE_COLUMNS = BASELINE_FEATURE_COLUMNS
 
 RAIN_EVENT_THRESHOLD_MM = 0.2
 RAIN_SPIKE_THRESHOLD_MM_5M = 5.0
-RAIN_HORIZON_BUCKETS = 12  # 12 * 5m = 1h
+BUCKET_MINUTES = 5
+DEFAULT_HORIZON_HOURS = 4
+SUPPORTED_PREDICTION_HORIZONS = (1, 4)
+
+
+def normalize_horizon_hours(horizon_hours: int) -> int:
+    out = int(horizon_hours)
+    if out <= 0:
+        raise ValueError("horizon_hours must be > 0")
+    return out
+
+
+def horizon_suffix(horizon_hours: int) -> str:
+    return f"{normalize_horizon_hours(horizon_hours)}h"
+
+
+def horizon_buckets(horizon_hours: int) -> int:
+    hours = normalize_horizon_hours(horizon_hours)
+    return (hours * 60) // BUCKET_MINUTES
+
+
+def rain_last_mm_col(horizon_hours: int) -> str:
+    return f"rain_last_{horizon_suffix(horizon_hours)}_mm"
+
+
+def rain_next_mm_col(horizon_hours: int) -> str:
+    return f"rain_next_{horizon_suffix(horizon_hours)}_mm"
+
+
+def rain_next_flag_col(horizon_hours: int) -> str:
+    return f"rain_next_{horizon_suffix(horizon_hours)}"
+
+
+def prediction_table_for_horizon(horizon_hours: int) -> str:
+    horizon = normalize_horizon_hours(horizon_hours)
+    if horizon == 1:
+        return "predictions_rain_1h"
+    if horizon == 4:
+        return "predictions_rain_4h"
+    raise ValueError(f"unsupported prediction-table horizon: {horizon_hours}")
 
 
 def parse_time(value: str) -> str:
@@ -232,6 +271,7 @@ def build_dataset(
     baro: pd.DataFrame,
     forecast: pd.DataFrame | None = None,
     rain_event_threshold_mm: float = RAIN_EVENT_THRESHOLD_MM,
+    horizon_hours: int = 1,
 ) -> pd.DataFrame:
     if ws90.empty:
         raise RuntimeError("no ws90 observations found")
@@ -261,12 +301,20 @@ def build_dataset(
     df["rain_inc"] = df["rain_inc_raw"].clip(lower=0)
     df["rain_spike_5m"] = df["rain_inc"] >= RAIN_SPIKE_THRESHOLD_MM_5M
 
-    window = RAIN_HORIZON_BUCKETS
-    df["rain_last_1h_mm"] = df["rain_inc"].rolling(window=window, min_periods=1).sum()
-    df["rain_next_1h_mm"] = df["rain_inc"].rolling(window=window, min_periods=1).sum().shift(-(window - 1))
-    df["rain_next_1h"] = df["rain_next_1h_mm"] >= rain_event_threshold_mm
+    windows: dict[int, int] = {
+        1: horizon_buckets(1),
+        normalize_horizon_hours(horizon_hours): horizon_buckets(horizon_hours),
+    }
+    for hours, window in windows.items():
+        rain_last_col = rain_last_mm_col(hours)
+        rain_next_mm = rain_next_mm_col(hours)
+        rain_next_flag = rain_next_flag_col(hours)
+        df[rain_last_col] = df["rain_inc"].rolling(window=window, min_periods=1).sum()
+        df[rain_next_mm] = df["rain_inc"].rolling(window=window, min_periods=1).sum().shift(-(window - 1))
+        df[rain_next_flag] = df[rain_next_mm] >= rain_event_threshold_mm
 
-    df["pressure_trend_1h"] = df["pressure_hpa"] - df["pressure_hpa"].shift(window)
+    window_1h = horizon_buckets(1)
+    df["pressure_trend_1h"] = df["pressure_hpa"] - df["pressure_hpa"].shift(window_1h)
 
     # Wind direction cyclical encoding.
     radians = np.deg2rad(df["wind_dir_deg"] % 360.0)
@@ -279,14 +327,14 @@ def build_dataset(
     df["wind_avg_lag_5m"] = df["wind_avg_m_s"].shift(1)
     df["pressure_lag_5m"] = df["pressure_hpa"].shift(1)
 
-    df["temp_roll_1h_mean"] = df["temperature_c"].rolling(window=window, min_periods=3).mean()
-    df["temp_roll_1h_std"] = df["temperature_c"].rolling(window=window, min_periods=3).std()
-    df["humidity_roll_1h_mean"] = df["humidity"].rolling(window=window, min_periods=3).mean()
-    df["humidity_roll_1h_std"] = df["humidity"].rolling(window=window, min_periods=3).std()
-    df["wind_avg_roll_1h_mean"] = df["wind_avg_m_s"].rolling(window=window, min_periods=3).mean()
-    df["wind_gust_roll_1h_max"] = df["wind_max_m_s"].rolling(window=window, min_periods=3).max()
-    df["pressure_roll_1h_mean"] = df["pressure_hpa"].rolling(window=window, min_periods=3).mean()
-    df["pressure_roll_1h_std"] = df["pressure_hpa"].rolling(window=window, min_periods=3).std()
+    df["temp_roll_1h_mean"] = df["temperature_c"].rolling(window=window_1h, min_periods=3).mean()
+    df["temp_roll_1h_std"] = df["temperature_c"].rolling(window=window_1h, min_periods=3).std()
+    df["humidity_roll_1h_mean"] = df["humidity"].rolling(window=window_1h, min_periods=3).mean()
+    df["humidity_roll_1h_std"] = df["humidity"].rolling(window=window_1h, min_periods=3).std()
+    df["wind_avg_roll_1h_mean"] = df["wind_avg_m_s"].rolling(window=window_1h, min_periods=3).mean()
+    df["wind_gust_roll_1h_max"] = df["wind_max_m_s"].rolling(window=window_1h, min_periods=3).max()
+    df["pressure_roll_1h_mean"] = df["pressure_hpa"].rolling(window=window_1h, min_periods=3).mean()
+    df["pressure_roll_1h_std"] = df["pressure_hpa"].rolling(window=window_1h, min_periods=3).std()
 
     # Calendar/seasonality features (UTC based).
     hour_of_day = df.index.hour + (df.index.minute / 60.0)
@@ -304,11 +352,16 @@ def build_dataset(
     return df
 
 
-def model_frame(df: pd.DataFrame, feature_cols: list[str] | None = None, require_target: bool = True) -> pd.DataFrame:
+def model_frame(
+    df: pd.DataFrame,
+    feature_cols: list[str] | None = None,
+    require_target: bool = True,
+    target_col: str | None = None,
+) -> pd.DataFrame:
     features = feature_cols or FEATURE_COLUMNS
     required = list(features)
     if require_target:
-        required.append("rain_next_1h")
+        required.append(target_col or rain_next_flag_col(1))
     out = df.dropna(subset=required).copy()
     return out.sort_index()