Air Quality Prediction ML Project

Table of Contents

1. Project Description
2. Dataset
3. Problem Statement
4. Solution Overview
5. Data Preparation
6. Exploratory Data Analysis (EDA)
7. Model Selection and Tuning
8. Project Structure
9. Installation
10. Usage
11. API Documentation
12. Deployment
13. Streamlit APP
14. Future Work
15. References

---

Project Description

This project aims to predict air pollution levels, specifically benzene concentration (C6H6(GT)), using environmental and sensor data.
A machine learning model is trained to provide fast predictions that can assist in monitoring and managing air quality.
The project includes data preprocessing, exploratory analysis, model training, hyperparameter tuning, and deployment as a REST API.

---

Dataset

• Source: UCI Machine Learning Repository – Air Quality Dataset
• Description: Contains hourly averaged responses from an array of chemical sensors embedded in an Air Quality Chemical Multisensor Device, along with meteorological data.
• Features:
  • Gas sensor measurements:
    • CO(GT): True hourly averaged concentration CO in mg/m^3
    • PT08.S1(CO): Tin Oxide sensor response (CO)
    • NMHC(GT): Non-methane hydrocarbons in microg/m^3
    • C6H6(GT): Benzene concentration in microg/m^3 (target)
    • PT08.S2(NMHC): Indium Oxide sensor response (NMHC)
    • NOx(GT): Nitrogen oxides concentration in ppb
    • PT08.S3(NOx): Tungsten Oxide sensor response (NOx)
    • NO2(GT): Nitrogen dioxide concentration in microg/m^3
    • PT08.S4(NO2): Tungsten Oxide sensor response (NO2)
    • PT08.S5(O3): Indium Oxide sensor response (O3)
  • Meteorological data:
    • T: Temperature in °C
    • RH: Relative Humidity (%)
    • AH: Absolute Humidity
  • Time-related features (engineered):
    • Year, Month, Day, Hour
    • Hour_sin, Hour_cos, Month_sin, Month_cos, Weekday_sin, Weekday_cos
• Target: C6H6(GT) – benzene concentration

---

Problem Statement

Air pollution is a critical public health issue.
Predicting benzene concentration helps:

• Alert communities about high pollution periods
• Enable authorities to take preventive measures
• Improve environmental monitoring

The goal is to create a predictive model that can be accessed via a REST API for real-time inference.

---

Solution Overview

• Model: RandomForestRegressor (with tuned hyperparameters)

• Pipeline:

  1. Data cleaning (handle missing values, duplicates, outliers)
  2. Feature engineering (sine/cosine transformations for cyclical features)
  3. Train/validation/test split
  4. Model training and hyperparameter tuning
  5. Model evaluation (RMSE, MAE, R²)
  6. Deployment as a FastAPI web service

• Input: Environmental sensor readings + time features

• Output: Predicted benzene concentration (C6H6(GT))

---

Data Preparation

• Missing values: Imputed using median or mean as appropriate
• Duplicates: Removed to ensure data quality
• Outliers: Treated using the IQR method for robust modeling
• Feature engineering: Added cyclical encodings for hour, month, weekday to capture periodicity

---

Exploratory Data Analysis (EDA)

• Correlation heatmaps: To identify relationships between features and target
• Time-series plots: Visualization of C6H6(GT) trends over time
• Distribution plots: For sensor readings and meteorological variables
• Feature importance: Extracted from RandomForest to interpret model decisions

---

Model Selection and Tuning

• Models evaluated:
  • Linear Regression
  • Decision Tree Regressor
  • Random Forest Regressor (best performance)
  • GradientBoostingRegressor
  • XGBRegressor
• Hyperparameter tuning: Performed using RandomizedSearchCV
• Evaluation metrics: RMSE, MAE, R² on validation and test sets
• Final model: Trained on combined train + validation data, saved as a serialized artifact (rf_model.pkl)

---

Installation

1. Clone the repository:

   git clone https://github.com/B-MEbrahim/Air-Quality-Prediction.git
   cd Air-Quality-Prediction

2. Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate  

3. Install dependencies:

   pip install -r requirements.txt

---

Usage

3. Run the API

Start the FastAPI server:

uvicorn predict:app --reload

4. Make Predictions

Send a POST request to the API endpoint with sensor and time features:

curl -X POST "http://127.0.0.1:8000/predict" \
     -H "Content-Type: application/json" \
     -d '{"CO_GT": 2.5, "PT08_S1_CO": 1200, "NMHC_GT": 150, "PT08_S2_NMHC": 900, "NOx_GT": 50, "PT08_S3_NOx": 600, "NO2_GT": 30, "PT08_S4_NO2": 800, "PT08_S5_O3": 1000, "T": 18.5, "RH": 45, "AH": 0.8, "Hour": 14, "Month": 5, "Weekday": 2}'

---

API Documentation

Endpoint

• POST /predict

Request Body

{
  "CO_GT": 2.3,
  "PT08_S1_CO": 1200,
  "PT08_S2_NMHC": 950,
  "NOx_GT": 140,
  "PT08_S3_NOx": 800,
  "NO2_GT": 110,
  "PT08_S4_NO2": 850,
  "PT08_S5_O3": 900,
  "T": 18.5,
  "RH": 45,
  "AH": 0.7,
  "Year": 2004,
  "Day": 123,
  "Hour_sin": 0.5,
  "Hour_cos": 0.8,
  "Month_sin": 0.1,
  "Month_cos": 0.99,
  "Weekday_sin": -0.3,
  "Weekday_cos": 0.95
}

Response

{
  "predicted_C6H6_GT": float
}

Example

{
  "predicted_C6H6_GT": 12.34
}

---

Deployment

• Deployed on Render: https://air-quality-prediction-28ic.onrender.com/

---

Streamlit Web App

To make the Air Quality Prediction model accessible and easy to use, an interactive Streamlit web application was built. This app allows users to input environmental conditions and sensor readings, sends the data to the deployed FastAPI backend (hosted on Render), and displays the predicted Benzene concentration (C6H6) along with an easy-to-understand air quality assessment.

• Try it out : https://air-quality-prediction01.streamlit.app/

Future Work

• Integrate real-time data ingestion from IoT sensors
• Add support for additional pollutants (e.g., NO2, CO)
• Implement model retraining pipeline for continuous improvement
• Enhance API with authentication and logging
• Build a dashboard for visualization and monitoring

---

References

• UCI Air Quality Dataset