Short-term Traffic Flow Forecasting Based On Deep Learning

With accelerations in the construction of smart cities in the recent years,numerous sensor devices have been deployed to various urban road networks for data collection of real-time traffic flow characteristics.By analyzing and processing this massive historical traffic data,it is possible to accurately forecast changes in traffic flow of various urban road networks.The predicted road traffic states can provide support to traffic managers as they create strategies,which thereby improve traffic safety,ease road congestions,and provide better travel experiences for the public.Short-time traffic flow forecasting is the basis of road traffic control and guidance,which is of great practical significance in road capacity improvements.Most of existing research work merely makes use of historical traffic data for prediction,without taking into account possible factors that affect traffic flow such as weather,holidays,points of interest,etc.In this paper,a hybrid model that combines multi-source heterogeneous data for traffic flow forecasting is proposed.First,the possible factors affecting road traffic flow are analyzed.Second,the raw heterogeneous data is prepared and pre-processed,and the dataset is divided in a certain proportion.Finally,a novel model that combines ensemble structures and transition matrices for forecasting short-term traffic flow is constructed.Considering the spatiotemporal characteristics of traffic data,the spatial correlation is modeled by employing road network distance.We leverage the ensemble structure to capture the periodicity,uncertainty,and nonlinearity of traffic flow,and generate the corresponding transition matrix to predict traffic volume.The experiment was performed using real-world traffic data from Xiamen,and the results indicated the following:(1)External factors such as weather,temperature,and holidays have little impact on traffic flow prediction.(2)Points of interest can improve the performance of the model for 30-minute traffic forecasting,but offers little help for 15-minute and 60-minute forecasting.(3)By comparing the prediction errors of variousmodels,the proposed model is more efficient and robust in terms of traffic forecasting.The introduction of the transition matrix can reduce the prediction error of the proposed model by up to 60%.