Ensemble Learning Traffic Model for Sofia: A Case Study

Abstract

Traffic models have gained much popularity in recent years, in the context of smart cities and urban planning, as well as environmental and health research. With the development of Machine Learning (ML) and Artificial Intelligence (AI) some limitations imposed by the traditional analytical, numerical and statistical methods have been overcome. The present paper shows a case study of traffic modeling with scarce reliable data. The approach we propose resorts on the advantages of ensemble learning using a large number of related features such as road and street categories, population density, functional analysis, space syntax, previous traffic measurements and models, etc. We use advanced regression models such as Random Forest, XGBoost, CatBoost etc., ranked according to the chosen evaluation metrics and stacked in a weighted ensemble for optimal fitting. After a series of consecutive data imputations we estimate the annual average daily traffic distribution in the street and road network of Sofia city and the metropolitan municipality for 2018 and 2022, and the NO2 levels for 2021 with accuracy resp. 78%, 74% and 92%, using AutoGluon and Scikit-Learn.