AQI Prediction Using Time Series Forecasting and MLflow Tracking

This project presents a full-stack machine learning pipeline to forecast Air Quality Index (AQI) using:

• Multi-step pollutant prediction via Echo State Networks (ESNs)
• Future AQI classification using XGBoost
• Experiment tracking and metric logging with MLflow

---

📌 Objective

• Predict AQI categories (0–4, mapped to Good → Very Poor) for next 6 time steps
• Use previously forecasted pollutant values to make these predictions
• Log and compare all results using MLflow

---

🔧 Pipeline Overview

✅ Step 1: Pollution Forecasting (ESN-based)

• Built individual Echo State Network models per pollutant (pm2.5, pm10, co, etc.)
• Trained models can be found inside "models" file
• Forecasted each pollutant 12 steps ahead using delay-embedded temporal inputs
• Visualized each forecast vs ground truth

✅ Step 2: AQI Classification (XGBoost-based)

• Used predicted pollutant values as inputs
• Trained multi-output XGBoost classifier to predict AQI levels at t+1 to t+6
• Labels categorized into AQI classes: Good (0) to Very Poor (4)

✅ Step 3: Experiment Tracking with MLflow

• Logged all training metrics (f1_macro, precision, recall, accuracy) for each step
• Saved classification reports and confusion matrices
• Auto-generated CSV summary of all MLflow runs

---

📁 Dataset

The dataset consists of pollution measurements for various Indian cities, with features:

date, aqi, co, no, no2, o3, so2, pm2_5, pm10, nh3

---

🧪 Future Work

• Address class imbalance using:
• SMOTE / RandomOverSampler
• Class-weighted loss

---

📦 MLflow Run Summary

• mlflow_summary_baseline.csv
• Includes metrics, params, tags for every model run.

---