Machine learning applied to traffic forecasting

Abstract

Linus Aronsson Aron Bengtsson Department of Computer Science and Engineering Chalmers University of Technology Abstract Forecasting future traffic situations is of great importance for improving existing Intelligent Transportation Systems (ITS). The forecasts allow the various ITS technologies to function proactively. This provides safer roads with less accidents and lower congestion levels. Forecasting urban road traffic is challenging as it often follows complex nonlinear temporal patterns. Traditional statistical forecasting techniques therefore struggle to achieve accurate predictions. In recent years the computing power and available historic traffic data has increased drastically. These are two of the main ingredients required for the field of Machine Learning (ML) to work. Many ML techniques have been shown to be capable of capturing nonlinear patterns in data, which makes it a good candidate for traffic forecasting. This thesis therefore explores various ML technologies and applies them to time series forecasting. Additionally, some traditional approaches to time series forecasting are evaluated as baselines for comparison. The experiments conducted used traffic data from central Gothenburg, which was manually gathered throughout the project. The conclusion was that the best ML techniques provided a higher forecasting accuracy for both short-term and long-term predictions. More advanced hyperparameter optimization and feature engineering would further improve the ML models.