haldwani-Rainfall-Nowcast-Forecast

🌧️ Haldwani Rainfall Nowcast & Forecast

A 5-model weighted ensemble for hourly rain prediction over Haldwani, Uttarakhand, India.
Built on open meteorological data from Open-Meteo, the system fuses Random Forest, Extra Trees, XGBoost, CatBoost, and LightGBM into a calibrated, uncertainty-aware forecast.

---

🗂️ Repository Contents

├── models/
│   ├── random_forest.pkl          # Scikit-learn RandomForestClassifier
│   ├── extra_trees.pkl            # Scikit-learn ExtraTreesClassifier
│   ├── xgboost.pkl                # XGBoost XGBClassifier
│   ├── catboost.pkl               # CatBoost CatBoostClassifier
│   ├── lightgbm.pkl               # LightGBM LGBMClassifier
│   ├── weighted_ensemble.pkl      # Full weighted fusion model
│   ├── feature_columns.pkl        # Ordered list of input features
│   ├── threshold.pkl              # Classification threshold (v1)
│   ├── production_config.pkl      # Per-model weights + threshold
│   ├── final_model_v2.pkl         # v2 single-model checkpoint
│   ├── final_features_v2.pkl      # v2 feature columns
│   └── final_threshold_v2.pkl     # v2 classification threshold
├── app.py                         # Streamlit forecast dashboard
├── predictor.py                   # RainPredictor inference class
├── feature_engineering.py         # Feature construction pipeline
├── weather_api.py                 # Open-Meteo API wrapper
├── config.py                      # Path configuration
├── requirements.txt               # Python dependencies
└── plots/                         # Diagnostic & analysis plots

---

🧠 Models

All five models are combined into a weighted ensemble via production_config.pkl, which stores each model's weight and the optimal classification threshold.

Uncertainty Estimation

The ensemble computes the unweighted standard deviation across all 5 model probabilities per forecast step. High std = models disagree = high uncertainty. This is used in the app to blend AI probability with Open-Meteo's POP (Probability of Precipitation) signal.

---

⚙️ Inference

Load & predict

import joblib
import pandas as pd
from predictor import RainPredictor
from weather_api import get_forecast_by_city, get_recent_observed

# Load the predictor (loads all models + config internally)
predictor = RainPredictor()

# Fetch live data from Open-Meteo
history_df  = get_recent_observed("Haldwani")
forecast_df = get_forecast_by_city("Haldwani")

# Run inference
results = predictor.predict(history_df, forecast_df)
print(results[["time", "rain_probability", "prediction_label"]].head(24))

Output columns

---

🚀 Run the Streamlit App

pip install -r requirements.txt
streamlit run app.py

The dashboard shows:

• Hourly rain forecast for the next 24 h
• Per-model probability breakdown
• Uncertainty-aware fusion with Open-Meteo POP
• Historical model diagnostics

---

📦 Dependencies

streamlit==1.47.1
pandas==2.3.1
numpy==2.3.2
scikit-learn
joblib==1.5.1
requests==2.32.4
plotly==6.2.0
python-dateutil==2.9.0.post0
xgboost
catboost
lightgbm

---

📍 Location

Haldwani, Uttarakhand, India
Lat: 29.2183° N, Lon: 79.5130° E
Timezone: Asia/Kolkata (IST, UTC+5:30)

Weather data sourced from Open-Meteo — free, no API key required.

---

📊 Diagnostics

Model evaluation plots are included under plots/:

• Per-model diagnostic charts (confusion matrix, ROC, calibration)
• Weighted ensemble vs fusion comparison
• Feature importance comparison across models
• Rolling accuracy, temporal F1, and weight sensitivity analysis

---

👤 Author

Aditya Suyal
huggingface.co/adityasuyal

---

📄 License

MIT