Traffic Flow and Congestion Prediction Using Machine Learning

Abstract

ABSTRACT Urban traffic congestion is a dynamic spatio-temporal phenomenon driven by recurring demand, incidents, weather, road capacity, signal operations, and heterogeneous vehicle behavior. Conventional traffic control is frequently reactive: a signal plan or route intervention is changed after congestion has already propagated. This paper develops an analytical and simulation-oriented machine-learning framework for short-term traffic-flow and congestion prediction using heterogeneous traffic observations. The proposed framework combines temporal traffic variables—flow, speed, occupancy, historical congestion, and time-of-day—with contextual variables such as rainfall, visibility, incidents, and calendar effects. Three supervised learning families, Random Forest, Gradient Boosting, and Long Short-Term Memory (LSTM), are compared against a persistence baseline. A rolling-window evaluation strategy is used to reduce temporal leakage, while congestion is represented through a congestion index derived from normalized speed and occupancy. Under the simulated experimental setting, the LSTM model produces the strongest short horizon performance because it captures sequential dependencies, while Gradient Boosting provides a competitive accuracy-to-computation trade-off for deployment on resource-constrained traffic-control systems. The analysis further shows that adding incident and weather features improves prediction during non-recurring congestion, although the benefit varies by forecast horizon. The findings support proactive signal timing, route guidance, incident response, and intelligent transportation-system decision support. The paper emphasizes that machine-learning prediction should complement, rather than replace, traffic engineering principles and should be evaluated continuously for drift, fairness across road segments, and robustness to missing sensor data.

KEYWORDS: Traffic flow prediction, congestion prediction, machine learning, LSTM, Random Forest, Gradient Boosting, intelligent transportation systems, spatio-temporal data, urban mobility