Skip to content

feature_comparison.md

Latest commit

 

History

History
156 lines (124 loc) · 5.13 KB

feature_comparison.md

File metadata and controls

156 lines (124 loc) · 5.13 KB
Raw

Feature Engineering Comparison

Current Implementation vs New feature_engineering.py

Tidal State Encoding MISMATCH ⚠️

Current (hysplit_forecasting.py line 138-148):

{
    'low': 0, 'slack low': 0,
    'rising': 1, 'flood': 1,
    'high': 2, 'slack high': 2,
    'falling': 3, 'ebb': 3,
}

New (feature_engineering.py line 97-98):

{
    'low': 0, 'ebb': 1, 'flood': 2, 'slack': 3, 'high': 4
}

Action Required: Update add_tidal_encoding() in hysplit_forecasting.py to match the new encoding

Feature Checklist

Features Currently Generated in modeldata_h2s ✓

Feature Source Function Present
Raw Weather
temperature_2m, wind_speed_10m, wind_direction_10m Merged from weather ✓
wind_gusts_10m, precipitation, relative_humidity_2m Merged from weather ✓
surface_pressure, cloud_cover, dewpoint_2m Merged from weather ✓
Raw Other
Flow (m^3/s)--Border Merged from streamflow ✓
tide_height, tidal_state Merged from tides ✓
H2S, h2s_measured From APCD data ✓
site_name, time Core columns ✓
Wind Features add_wind_features()
wind_direction_sin, wind_direction_cos ✓ ✓
wind_direction_categorical ✓ ✓
wind_direction_categorical_encoded ✓ ✓
wind_speed_10m_avg_2h/3h/4h ✓ ✓
wind_gusts_10m_max_2h/3h/4h ✓ ✓
wind_temp_interaction ✓ ✓
humidity_temp_interaction ✓ ✓
Tidal add_tidal_encoding()
tidal_state_encoded ✓ ⚠️ WRONG MAPPING
Time Features add_inference_features()
hour_sin, hour_cos ✓ ✓
month_sin, month_cos ✓ ✓
is_night ✓ ✓
day_night add_day_night() ✓
Flow Features add_inference_features()
flow_log, flow_low, flow_high ✓ ✓
Other
source_regime add_inference_features() ✓
stable_atm add_inference_features() ✓
H2S Lag Features add_h2s_lag_features()
h2s_lag_1h, h2s_lag_3h, h2s_lag_6h ✓ ✓
h2s_rolling_6h, h2s_rolling_24h ✓ ✓
flow_lag_6h, flow_rolling_24h ✓ ✓
SBIWTP Features add_sbiwtp_features()
sbiwtp_flow_mgd, sbiwtp_anomaly, sbiwtp_deficit ✓ ✓
sbiwtp_flow_x_temp, sbiwtp_hourly_mgd, sbiwtp_sli ✓ ✓
Risk
h2s_risk Lines 684 ✓ (log-logistic formula)

Features in model_forecast ✓

All same as modeldata_h2s EXCEPT:

  • ✗ No H2S column
  • ✗ No h2s_measured column
  • ⚠️ h2s_lag_* and flow_lag_* features are filled by forecast_features.engineer_features() from utils/forecast_features.py
  • ✗ No h2s_risk (informational only, not needed for forecast)

Issues Found

1. ⚠️ CRITICAL: Tidal State Encoding Mismatch

The current implementation uses a different mapping than feature_engineering.py.

Impact: If the model was trained with the new encoding but inference uses the old one, predictions will be wrong.

Fix Required:

# In hysplit_forecasting.py, line 138-148, change to:
def add_tidal_encoding(tidal_df):
    """Add tidal_state_encoded column based on tidal_state."""
    tidal_mapping = {
        'low': 0,
        'ebb': 1,
        'flood': 2,
        'slack': 3,
        'high': 4,
    }
    if 'tidal_state' in tidal_df.columns:
        tidal_df['tidal_state_encoded'] = tidal_df['tidal_state'].map(tidal_mapping).fillna(1).astype(int)
    else:
        tidal_df['tidal_state_encoded'] = 1
    return tidal_df

2. ✓ H2S Risk Calculation Different

Current (hysplit_forecasting.py line 684):

matched_df['h2s_risk'] = matched_df['H2S'].pow(1.23) / (matched_df['H2S'].pow(1.23) + 5**1.23)

New (feature_engineering.py lines 166-169):

df['h2s_risk'] = 'GREEN'
df.loc[df['H2S'] > 5, 'h2s_risk'] = 'YELLOW'
df.loc[df['H2S'] > 10, 'h2s_risk'] = 'ORANGE'
df.loc[df['H2S'] > 30, 'h2s_risk'] = 'RED'

Impact: Current implementation uses a continuous risk score (0-1), new uses categorical. These serve different purposes.

Recommendation: Keep both:

  • Current h2s_risk (continuous) for numerical analysis
  • Add new categorical version as h2s_risk_category if needed

Recommendations

Immediate Actions

  1. Fix tidal_state_encoded mapping to match feature_engineering.py
  2. Verify which encoding was used during model training
  3. Test model predictions after fixing encoding to ensure accuracy

Optional Improvements

  1. Consolidate feature engineering - Replace current functions with calls to feature_engineering.py to ensure consistency
  2. Add categorical h2s_risk if needed for reporting
  3. Document which tidal states actually appear in your data

Feature Count Summary

Dataset Current Features Required by feature_engineering.py
modeldata_h2s ~55+ columns 53 (ENGINEERED_COLUMNS_OBS)
model_forecast ~55+ columns Same minus H2S, h2s_measured, h2s_risk

Status: ✓ All required features present, but ⚠️ tidal encoding mismatch must be fixed