🚦 Traffic Flow Forecasting using Spatial CNNs

This project implements a Spatial Convolutional Neural Network (CNN) to forecast urban traffic flow.
It combines grid-based sensor mapping, synthetic/real traffic data, deep learning forecasting, and unsupervised clustering for traffic pattern discovery.

---

⚡ Features

• Convert sensor distance matrix → 2D grid mapping (via MDS)
• Create traffic heatmaps sequences
• Train & tune a 3D CNN with Keras Tuner
• Save best model (models/cnn_model.keras)
• Visualize predicted vs actual heatmaps
• Apply KMeans clustering for unsupervised traffic pattern discovery
• Reduce with PCA for visualization

---

🛠️ Installation

1. Clone the repo

git clone https://github.com/hariravi-ds/Traffic-Flow-Forecasting-Using-Spatial-CNNs.git

cd Traffic-Flow-Forecasting-Using-Spatial-CNNs

2. Create virtual environment (recommended, macOS/Linux)

python3 -m venv .venv

source .venv/bin/activate

3. Install dependencies

pip install --upgrade pip

pip install -r requirements.txt

▶️ Running the Pipeline

python main.py

This will: Train + tune CNN

Save best model → models/cnn_model.keras

Save results → results/ folder:

pred_vs_actual.png (forecast comparison)

cluster_centers.png (unsupervised clusters)

pca_clusters.png (PCA scatter of clusters)

📦 Requirements

See requirements.txt.

Key libraries: TensorFlow / Keras

Keras-Tuner

scikit-learn

numpy, pandas

matplotlib, seaborn

🔮 Next Steps

Replace synthetic generator with real-world datasets (e.g. METR-LA, PeMS)

Add spatio-temporal GNNs for comparison

Experiment with attention-based forecasting