Abstract: Traffic flow information prediction is an important foundation for traffic guidance management of intelligent transportation systems, and provides reliable data support and scientific decision-making basis for urban traffic management planning. Since the traffic flow data is real-time updated incremental flow data, each time the historical data set is updated, the prediction model needs to be rebuilt, which consumes a lot of computing resources and running time. Therefore, this paper proposes an improved online sequential extreme learning machine for traffic flow prediction (IOS-ELM), which constructs the enhanced feature mapping relationship of the newly input data, generates the dynamic update feature representation space of the traffic flow, and realizes the dynamic update of the short-term traffic flow prediction model. Finally, the model is evaluated on the real-world traffic flow data of Yuanda 1st Road in Changsha, China. The experimental results show that the IOS-ELM model exceeds other baselines prediction models (MLP, ANN, ELM, OS-ELM) in the prediction performance of NRMSE and MAPE. Meanwhile, the computation prediction of IOS-ELM is less time-consuming, which can ensure a certain real-time performance and meet the needs of real-time and accurate prediction of urban road traffic flow.

Key words: traffic flow prediction, extreme learning machine, intelligent transportation system