🇮🇳 India Air Quality Prediction System
AI-powered Artificial Neural Network for forecasting air pollution in Indian tier-2 cities.
This project implements a Bidirectional LSTM with Attention neural network to predict PM2.5, PM10, NO2, SO2, O3, CO, and AQI for the next 6 hours based on 24 hours of historical data.
🎯 Target Cities
| City | Profile | Typical AQI |
|---|---|---|
| Lucknow | Industrial + vehicular | Moderate-Poor |
| Noida | Delhi-NCR corridor, severe winters | Poor-Very Poor |
| Bengaluru | Traffic-dominated, moderate | Satisfactory-Moderate |
| Jaipur | Dust + industrial | Moderate |
| Srinagar | Low-moderate, seasonal | Good-Satisfactory |
| Mumbai | Coastal, monsoon washout | Satisfactory-Moderate |
🧠 Model Architecture
Input (24h × 93 features)
↓
Bidirectional LSTM (64 hidden, 1 layer)
↓
Attention Mechanism
↓
MLP Decoder (256 → 128 → 42)
↓
Output (6h × 7 targets: PM2.5, PM10, NO2, SO2, O3, CO, AQI)
Parameters: ~161K
Framework: PyTorch
📊 Performance Metrics
| Metric | PM2.5 | PM10 | NO2 | SO2 | O3 | CO | AQI |
|---|---|---|---|---|---|---|---|
| RMSE | 0.25 | 0.25 | 0.38 | 0.15 | 0.89 | 0.23 | 18.71 |
| MAE | 0.20 | 0.20 | 0.34 | 0.12 | 0.89 | 0.20 | 15.13 |
| R² | 0.84 |
Metrics computed on held-out test set with inverse-transformed (original scale) values.
🗂️ Dataset
The dataset was generated using realistic pollution profiles based on CPCB (Central Pollution Control Board) documented patterns for Indian cities:
- Period: 3 years of hourly data (2023-2025)
- Records: 157,680 per city
- Features: Pollutants, meteorology, temporal cyclical encodings, lag features, rolling statistics
- Source: Synthetic, parameterized from published Indian air quality literature
Dataset Structure
data/india_air_quality_synthetic.csv
├── datetime, city
├── pm25, pm10, no2, so2, o3, co
├── temperature, humidity, wind_speed
├── hour, day_of_week, month, day_of_year
├── aqi, aqi_category
🚀 Usage
Training the Model
pip install -r requirements.txt
python generate_synthetic_data.py # Create dataset
python preprocess.py # Feature engineering
python train.py # Train ANN
Running the Prediction App
python app.py
Python API
import torch
from model import get_model
model = get_model(model_type='lstm_attn', input_dim=93)
model.load_state_dict(torch.load('models/best_model.pt')['model_state_dict'])
# Input: (batch, 24h, 93_features) -> Output: (batch, 6h, 7_targets)
📁 Repository Structure
├── data/ # Raw and processed data
├── processed/ # Scaled sequences for training
├── models/ # Trained checkpoints
├── model.py # ANN architectures
├── train.py # Training script
├── preprocess.py # Data preprocessing pipeline
├── generate_synthetic_data.py # Dataset generator
├── app.py # Gradio prediction interface
└── requirements.txt # Dependencies
🔗 References
- CPCB Air Quality Portal: https://airquality.cpcb.gov.in
- CPCB AQI Calculation Methodology (CPCB, 2014)
- AirPhyNet: Physics-Guided Neural Networks for Air Quality Prediction (Hettige et al., 2024)
- AirCast: Multi-Variable Air Pollution Forecasting (Nedungadi et al., 2025)
⚖️ Disclaimer
The current dataset is synthetic and parameterized from published pollution profiles for demonstration and model development. For production deployment, integrate with real-time CPCB CAAQMS data feeds or OpenAQ API with proper API keys.
📜 License
Apache-2.0