Prediction of Urban Traffic Flow Based on Generative Neural Network Model

Prediction of traffic congestions caused by excessive rush-hour traffic, accidents or road works, is an important issue of intelligent traffic management systems. The article will present a generative model of movement in a network of streets and intersections that enables to predict traffic congestions of an accidental or deterministic character. The model has a modular, hierarchical character. At the lowest level the model is microscopic and depicts the movement of vehicles, public transport and pedestrians at intersections and between them. It allows comparing real measurements with simulated data and attach data obtained from the learned neural network. At its higher level, the model reflects the traffic in a selected sub-area of the network of intersections. On this level the model is macroscopic, taking into account the parameters of vehicle streams. On both levels, the model is based on a generative deep neural network. A multi-layered neural network was chosen from many available deep architectures. In order to train this network the following methods were used: algorithms prepared for Restricted Boltzmann Machine and Deep Believe Network, as well as neighbourhood network. These architectures and methods adapt to the real dependencies of traffic much better and more accurately than traditional structures and methods of learning. Thanks to the mentioned solutions it is possible to eliminate damaged or incorrect data from data processing. One of the aims of the proposed model is to predict changes in the intensity traffic and traffic congestions in short-term forecasts, also in cases that have not yet occurred.