Traffic Volume Forecasting using ARIMA & Temporal Fusion Transformer (TFT)

This project demonstrates an end-to-end traffic forecasting system using:

• ARIMA – classical statistical model
• TFT (Temporal Fusion Transformer) – state-of-the-art deep learning model

Click here to see the dashboard

It includes:

• Full preprocessing pipeline
• Baseline model (ARIMA)
• Advanced deep learning model (TFT)
• Performance comparison
• Streamlit web dashboard

📘 Dataset

• Source: UCI Machine Learning Repository
• Name: Metro Interstate Traffic Volume
• Link: https://archive.ics.uci.edu/dataset/492/metro+interstate+traffic+volume
• Total Rows: 48,205
• Frequency: Hourly traffic volume
• Target Variable: traffic_volume

This dataset contains weather, time, holidays, and hourly traffic flow measurements.

🎯 Project Objective

To build a robust multi-step (24-hour) traffic forecasting application and demonstrate why TFT significantly outperforms ARIMA for real-world traffic prediction.

🛠 Tech Stack

Machine Learning

• ARIMA
• Temporal Fusion Transformer (PyTorch Forecasting)
• PyTorch Lightning

Data Engineering

• Pandas, Numpy
• Feature engineering (lags, rolling windows, embeddings)

Deployment

• Streamlit web app
• Model checkpoint loading (no retraining required)

🧹 Data Preprocessing

Key feature engineering steps:

• Parsed timestamps → extracted:
  hour, weekday, weekend, month, rush_hour
• Encoded categorical features:
  weather_main, weather_description, holiday
• Added lag features:
  • lag_1 (previous hour)
  • lag_24 (previous day same hour)
• Added rolling averages:
  • roll_3 (3-hour mean)
  • roll_24 (24-hour mean)
• Train-val-test split:
  • Train: 2012–2016
  • Val: 2017
  • Test: 2018

📈 Models Implemented

1. ARIMA (Baseline)

• Univariate
• Fast to train
• Struggles with non-linear seasonal behavior

ARIMA RMSE: ≈ 1989

---

2. Temporal Fusion Transformer (TFT)

A powerful attention-based architecture for multi-horizon forecasting.

Uses:

• Covariates (weather, time, rolling features)
• Variable selection networks
• Static & dynamic embeddings
• Multi-head temporal attention

TFT Performance :

• Overall RMSE ≈ 287
• t+1 RMSE ≈ 169
• t+6 RMSE ≈ 457
• t+24 RMSE ≈ 63

This represents over 85% performance improvement compared to ARIMA.

Why This Project Matters (Real-World Use Cases)

Traffic forecasting is critical for:

Smart Traffic Signals

Adaptive signal control reduces congestion.

Route Optimization

Used by Google Maps & Waze for accurate ETA prediction.

Public Transport Scheduling

Demand-aware bus/train scheduling.

Emergency Services

Optimized routing for ambulances, fire services, and police.

Urban Planning

Helps identify bottlenecks, plan flyovers, signal placements.

Pollution Reduction

Less idling → lower emissions → healthier cities.

This project replicates the core logic used in Intelligent Transportation Systems (ITS) and Smart City infrastructure.

🖥 Streamlit Dashboard

The Streamlit app includes:

• ARIMA forecast visualization
• TFT multi-horizon forecast
• Model performance metrics
• Comparison plots
• Dataset overview
• GitHub link and modern UI layout

Run locally:

git clone https://github.com/a-anuj/tft-traffic-volume-analysis.git
cd tft-traffic-volume-analysis
streamlit run app.py

Results Summary

• TFT outperforms ARIMA across all forecast horizons
• Learns complex seasonal and weather-driven traffic patterns
• Shows strong stability and accuracy for long-term predictions
• Suitable for real-world deployment in smart transportation systems
• Produces reliable and consistent 24-hour forecasts