A novel approach to traffic modelling based on road parameters, weather conditions and GPS data using feedforward neural networks

Abstract

This article presents the development of a estimation model using an artificial neural network to estimate the traffic factor parameter, which reflects changes in travel time for a single transport connection. The modeling of traffic is highly challenging due to the complex and non-linear nature of road systems, the interactions between diverse road users, and the ripple effects of congestion. The research problem addressed is the need to consider a wide range of factors that contribute to congestion, as identified in the literature. To address this challenge, a universal Feedforward Neural Network (FNN) was designed, specifically the Multilayer Perceptron (MLP) 38-10-1 model. The input layer neurons of the MLP receive a vector of input values representing various factors such as road types, technical properties, time of day, days of the week, incidents, weather conditions, and population density. The model was trained using a dataset consisting of 41,945 records of fully completed input variables, extracted from a larger dataset of 300,000 travel time measurements collected from a real transport network mapped on a macro scale (Mazovia, Poland). During the research, the model achieved a satisfactory level of Mean Absolute Percentage Error (MAPE) for the test set (9.12%) and a Normalized Root Mean Squared Error (NRMSE) below 0.02, indicating good estimation performance. Furthermore, this work proposes specialized models with the structure of MLP 34-10-1 for individual types of roads, which also demonstrated satisfactory estimation abilities. These models have practical applications in traffic management and planning. Overall, this research addresses the research problem of modeling factors for predicting travel time changes using an artificial neural network approach. The developed models provide accurate estimations and offer potential applications in transportation systems.

Topics & Concepts