PM2.5 Air Quality Forecasting (Lucknow)

Cloud-Ready Microservice: Predictive Modeling and Diagnostic Analysis of Urban Air Quality.

This repository contains a containerized machine learning pipeline engineered to forecast daily PM2.5 (particulate matter ≤ 2.5 µm) concentrations in Lucknow, India. The project focuses on feature engineering (Temporal Lags) and model comparison to provide actionable public health recommendations.

---

Project Architecture

The pipeline follows a structured data science workflow:

1. Data Ingestion: Processing daily pollutant datasets (CPCB standards) with forward/backward fill persistence for gap handling.
2. Feature Engineering:
   • Lags: pm25_lag1 and pm25_lag7 to capture daily and weekly cycles.
   • Rolling Window: 7-day moving averages to smooth volatility.
3. Modeling: Comparative analysis between Linear Regression (Baseline) and Random Forest Regressor (Non-linear ensemble).

---

Mathematical Evaluation

We utilize four formal definitions to evaluate model fidelity and feature construction:

1. Mean Absolute Error (MAE): Represents the average magnitude of errors without considering their direction. $$MAE = \frac{1}{n} \sum_{i=1}^{n} |y_i - \hat{y}_i|$$

2. Root Mean Square Error (RMSE): Penalizes larger forecasting errors, critical for detecting dangerous pollution spikes. $$RMSE = \sqrt{\frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2}$$

3. Coefficient of Determination ($R^2$): Measures the proportion of variance in PM2.5 levels explained by features. $$R^2 = 1 - \frac{\sum (y_i - \hat{y}_i)^2}{\sum (y_i - \bar{y})^2}$$

4. Rolling Feature Calculation: Capturing the short-term trend ($t$) over a window ($k=7$ days). $$\mu_t = \frac{1}{k} \sum_{i=0}^{k-1} x_{t-i}$$

---

Key Insights

• Performance: Random Forest outperformed Linear Regression by effectively capturing the sharp non-linear peaks in PM2.5 levels.
• Feature Importance: Time-based features (Rolling Means) and CO/SO2 concentrations were identified as the highest predictors of localized spikes.

---

Recommendations Framework

The system classifies forecasts into EPA-aligned categories to generate automated health advisories:

Category	PM2.5 Range (µg/m³)	Recommendation
Good	0 – 12.0	Air quality is satisfactory. Enjoy outdoor activities freely.
Moderate	12.1 – 35.4	Acceptable air quality. Sensitive individuals should limit prolonged exertion.
Unhealthy for Sensitive Groups	35.5 – 55.4	Sensitive groups should reduce outdoor activity. Consider wearing a mask.
Unhealthy	55.5 – 150.4	Everyone may experience health effects. Limit outdoor activities. Use purifiers.
Very Unhealthy	150.5 – 250.4	Health alert: serious effects possible. Avoid outdoor activity. Keep windows closed.
Hazardous	> 250.4	Emergency conditions. Stay indoors with filtered air. Follow local advisories.

---

System Architecture & Data Flow

This microservice is designed to ingest local meteorology data and execute inference within an isolated container runtime.

graph TD
    A[(Proprietary CSV Data)] -->|Mounted Volume / ENV path| B
    
    subgraph Docker Container [Containerized Microservice: python:3.10-slim]
        B[Data Ingestion & Preprocessing]
        C[Feature Engineering: Lags & Rolling Means]
        D[Scikit-Learn: Random Forest Inference]
        
        B --> C
        C --> D
    end
    
    D -->|Stdout / Logs| E[Predicted PM2.5 Levels]
    D -->|Logic Rule Engine| F[Automated Health Advisories]

Loading

---

Data Requirements (Proprietary)

Note: The raw meteorological and pollution dataset used to train this model is proprietary and not included in this public repository.

To run this pipeline with your own data, provide a CSV file at ./data/ML_Lucknow.csv with the following schema:

• date: (YYYY-MM-DD format)
• pm25: Target variable (float)
• co, so2, no2, o3: Chemical features (float)
• temp, humidity: Meteorological features (float)

---

Deployment & Installation

This forecasting pipeline is fully containerized for reproducible execution across different environments. Ensure your dataset is placed at ./data/ML_Lucknow.csv or specify a custom path via the DATA_PATH environment variable.

Option A: Run via Docker (Recommended)

git clone https://github.com/alfayezahmad/ideal-sniffle.git
cd ideal-sniffle

# 1. Build the container image
docker build -t ideal-sniffle .

# 2. Execute the model
docker run ideal-sniffle

Option B: Local Development

git clone https://github.com/alfayezahmad/ideal-sniffle.git
cd ideal-sniffle

# Install strict dependencies
pip install -r requirements.txt

# Run the pipeline
python main.py

---

Roadmap & Future Work

Currently, this pipeline operates as a standalone containerized batch-inference script. The next phase of R&D focuses on evolving it into a fully distributed, cloud-native microservice:

• REST API Integration: Wrap the inference engine in FastAPI to serve real-time predictions via HTTP endpoints rather than standard output.
• Continuous Integration (CI/CD): Implement GitHub Actions to automate Docker image builds and testing upon new commits.
• Kubernetes Orchestration: Develop Helm charts/K8s manifests to deploy and scale the containerized API within a distributed cluster.
• Real-Time Data Ingestion: Transition from static CSV files to a live MQTT or Apache Kafka stream connected to physical IoT air quality sensors.

---

License

Distributed under the MIT License. Author: Alfayez Ahmad | Copyright: © 2026