🌍 Air Quality Time Series Forecasting

This project focuses on forecasting PM2.5 air quality readings using a variety of time series modeling techniques, progressing from data wrangling to advanced ARMA forecasting.
Each stage introduces new analytical and modeling concepts, forming a complete end-to-end time series pipeline.

---

📄 Overview

Air pollution, especially PM2.5 (Particulate Matter < 2.5µm), is a key indicator of environmental health.
This project aims to predict PM2.5 concentrations over time using data collected from the Nairobi Air Quality Database.

We start by cleaning and transforming data, then move through Linear Regression, Autoregressive, and ARMA models to improve forecasting accuracy.

---

🧮 Dataset

• Source:open.africa/dataset open.Africa database
• Location: Nairobi, Kenya
• Features: Timestamped PM2.5 air quality readings (hourly frequency)
• Cleaning: Missing values handled via forward fill; timezone localized to Africa/Nairobi

---

⚙️ Workflow

1. Data Extraction & Wrangling:
   Downloaded the data in CSV form and transformed raw air quality data into a clean, structured format.

2. Feature Engineering:
   Created lag-based and time-derived features to capture temporal patterns and dependencies.

3. Modeling & Forecasting:
   Built multiple models — Linear Regression, Autoregressive (AR), and ARMA — to progressively enhance prediction accuracy.

4. Evaluation & Visualization:
   Compared model outputs using MAE and plotted actual vs. predicted PM2.5 levels for interpretability.

---

🧩 Models Used

📈 Linear Regression with Time Series Data

📘 Notebook: Linear_regression.ipynb

• Built a Linear Regression model on time-based features.
• Modeled PM2.5 as a function of historical timestamps.
• Served as a baseline model for comparison with time-dependent methods.

---

🔁 Autoregressive (AR) Model

📘 Notebook: AutoReg.ipynb

• Trained an AR(p) model to predict PM2.5 based on previous hourly readings.
• Used lag order = 26, chosen based on lowest Mean Absolute Error (MAE).
• Implemented walk-forward validation to simulate real-time forecasting.

---

⚙️ ARMA Model & Hyperparameter Tuning

📘 Notebook: ARMA model.ipynb

• Combined autoregressive (AR) and moving average (MA) terms into an ARMA(p, q) model.
• Tuned parameters p and q via grid search using MAE-based heatmaps.
• Identified the optimal configuration that balanced accuracy and computational efficiency.

---

📊 Evaluation

• Metric Used: Mean Absolute Error (MAE)
• Compared model predictions on both training and test data.
• Best performance achieved by the ARMA(8, 0, 1) configuration during walk-forward validation.
• Evaluation confirmed model stability and ability to adapt to real-time sequences.

---

📈 Visualization

• Used Plotly Express to plot actual vs. predicted PM2.5 readings.
• Created interactive time series charts to highlight model performance.
• Visualizations helped identify periods where predictions aligned or deviated from observed data.

---

🏁 Results and Insights

• Cleaned and modeled over one month of hourly air quality readings.
• Demonstrated improvements from baseline (Linear Regression) to AR/ARMA models.
• Walk-forward validation proved the model’s adaptability in sequential forecasting.
• Interactive visualizations made trend analysis intuitive and engaging.

---

💡 Conclusion

This project demonstrates a complete time series forecasting workflow — from data wrangling to predictive modeling and visualization.
It highlights core skills in:

• Time series analysis
• Model evaluation and tuning
• Real-world data handling
• Interactive visualization using Plotly

The techniques and methodology here are directly applicable to data analytics, environmental monitoring, and predictive modeling projects.

---

🧠 Tech Stack

• Languages: Python
• Libraries: Pandas, Statsmodels, Scikit-learn, Plotly, Seaborn, Matplotlib
• Tools: Jupyter Notebook

---

✨ Author

Pushkin Kumar
Data Analyst | Data Engineer
Passionate about building modern data pipelines and predictive models.