This paper investigates the effectiveness of multi-variate deep learning models for traffic flow prediction with road incidents. Multiple features of the data are considered in the analysis including traffic features, road incident features and cyclical features. Three multivariate deep learning models based on the stacked LSTM network, the CNN LSTM network, and the Autoencoders-LSTM network are developed for long-short term traffic forecasting under road incidents. The results obtained from the analysis are then compared to determine the best suitable approach.

This paper focuses on the microsimulation of traffic flow on freeways with on-ramps, off-ramps and bottlenecks. The road network model with many vehicle routes is constructed based on the detector flow data recorded every minute on each freeway lane. A comparative study evaluates the effectiveness of the proposed traffic flow model. It is noted that the model can successfully capture the essential features of traffic flow on the freeway. With the calibrated model, various numerical experiments are then conducted to investigate the impact of ramp metering (RM) and variable speed limit (VSL) on average waiting time, average travel time, average traffic speed, and the total number of running vehicles over certain period of time on the road network. The results indicate that the RM and VSL control can significantly improve traffic flow on the freeway.