Karachi AQI Forecasting System

Overview

Karachi AQI Forecasting System is an end-to-end machine learning project developed during my Data Science internship at 10Pearls SHINE. The system was designed to forecast Karachi's Air Quality Index (AQI) across 24-hour, 48-hour, and 72-hour horizons using continuously updated environmental data and automated ML pipelines.

Problem

Air quality can change considerably over short periods, making current AQI readings alone insufficient for understanding upcoming conditions. The objective was to build a forecasting pipeline capable of transforming continuously collected environmental data into future AQI predictions while also supporting model experimentation, evaluation, and monitoring.

Approach

I developed an automated feature pipeline that collected and processed hourly environmental data and stored engineered features in the Hopsworks Feature Store.

Multiple forecasting approaches were explored, including:

• Ridge Regression
• Random Forest
• XGBoost
• LSTM

Models were evaluated across multiple forecasting horizons to understand how predictive performance changed as the forecast window increased.

Hopsworks was used for feature management and model registry functionality, while GitHub Actions automated recurring pipeline execution.

ML Pipeline

The system followed an end-to-end workflow:

1. Collect updated environmental and AQI data.
2. Clean and transform incoming observations.
3. Engineer forecasting features.
4. Store features in the Hopsworks Feature Store.
5. Train and evaluate forecasting models.
6. Register model artifacts and results.
7. Generate AQI forecasts for multiple horizons.
8. Run the data pipeline automatically using GitHub Actions.

Model Evaluation

The project compared several machine learning and deep learning approaches rather than relying on a single forecasting model.

During one stable 24-hour forecasting run, the Ridge Regression model achieved:

• RMSE: 7.74
• MAE: 5.60
• R²: 0.297

Performance was monitored as the available dataset evolved, which also highlighted practical challenges associated with limited samples, changing data distributions, and longer forecasting horizons.

Engineering

Beyond model training, the project focused on building a reproducible ML workflow.

Hopsworks provided centralized feature storage and model management, while automated GitHub Actions workflows allowed the feature pipeline to update without requiring manual execution.

This made the project an exercise not only in predictive modeling, but also in feature engineering, automation, model management, and production-oriented ML workflow design.

What I Learned

The project demonstrated how real-world forecasting performance can change as new data becomes available. It also highlighted the importance of monitoring data coverage, avoiding conclusions based on very small evaluation samples, and treating the surrounding data pipeline as an important part of an ML system rather than focusing only on model accuracy.