Flowcast — Melbourne Traffic Intelligence Platform

What It Does

Flowcast is a traffic intelligence platform for Melbourne, Australia. It ingests 131 million rows of Victorian SCATS traffic signal data — 5.3 GB compressed across 4,741 intersections — to train gradient-boosted forecasting models and present results through an interactive WebGL dashboard with live map rendering.

How It Works

A resumable Python pipeline loads raw SCATS ZIP files into DuckDB using wide-format storage (96 columns per row) to avoid materializing 12.6 billion long-format rows. KMeans clustering groups intersections into 8 behavioral profiles based on 24-hour traffic shape. A HistGradientBoostingRegressor (500 trees) forecasts daily vehicle counts at 1, 7, 14, and 28-day horizons with recursive multi-step prediction and volatility-aware confidence intervals. The FastAPI backend serves forecasts, clusters, and correlation data to a React + MapLibre frontend rendered with WebGL.

Key Features

• WebGL map — 4,449 geocoded intersections with volume, cluster, and region color modes
• Multi-horizon forecasting — 1/7/14/28-day predictions at 3.9% MAPE with confidence bands
• Behavioral clustering — 8 intersection profiles visualized with radar and scatter charts
• Corridor detection — Pearson correlation + cosine similarity identifies linked site pairs with time-lag estimation
• Resilient pipeline — Idempotent ingestion with manifest-based tracking and quality scoring