Real-time congestion prediction for urban arterials using adaptive data-driven methods

Congestion prediction can support traffic departments to take effective traffic management measures, and it aids users to adopt smarter trip strategies, including route and departure time selection. This paper proposes an adaptive data-driven real-time congestion prediction method. The method framework includes a traffic pattern recognition algorithm based on the adaptive K-means clustering, a two-dimensional speed prediction model and an adaptive threshold calibration method. After the principal component analysis, the adaptive K-means cluster algorithm is conducted to obtain different traffic patterns. Congestion recognition is realized with an adaptive threshold calibration method and congestion prediction is then raised according to different traffic patterns. Parameter calibration and model evaluation are carried out on the proposed method using floating car travel speed data. The results show that the adaptive K-means cluster recognition algorithm can better recognize the traffic patterns than Gaussian Mixture model, and the proposed congestion prediction method has better real-time performance, accuracy and robustness than Autoregressive Integrated Moving Average model and Kalman Filtering method.