Deep-PRESIMM: Integrating Deep Learning with Microsimulation for Traffic Prediction

Abstract

This paper presents a proactive model and tool for urban traffic analysis and management that integrates deep learning for traffic parameter prediction with traffic microsimulation, providing traffic analysts with the ability to visualise the traffic network state ahead of time, generate traffic control measures, and analyse the consequences of the applied traffic control measure(s). The model adopts an integrated assess-forecast-simulate approach in which traffic flow characteristics are applied on deep Convolutional Neural Network Long Short-Term Memory (CNN-LSTM) stacked autoencoders to forecast traffic flow and speed, which are subsequently passed on to a traffic microsimulation tool – Simulation of Urban Mobility (SUMO) – where the predicted parameters are used to generate a traffic future state simulation. The model is evaluated using sensor-collected historical traffic and weather data from Stretford, Greater Manchester in the United Kingdom. The results show the promise of the proposed model when applied towards urban traffic management and control.