Urban traffic flow online prediction based on multi‐component attention mechanism

Abstract

Traffic flow prediction is regarded as an important concept used in traffic planning, traffic design, and traffic management. In this study, the authors propose a multi-component attention (MCA) method for traffic flow prediction, which may jointly and adaptively understand components of long-term trends, seasons, and traffic flow residuals that result from multi-dimensional decomposition. According to the highly non-linear nature of traffic flow, the proposed module consists of a one-dimensional convolutional neural network, a bidirectional long short-term memory, and a bidirectional mechanism with an attention mechanism. The former captures local trend characteristics of residuals, while the latter captures trends and seasonal time adjustments. Due to the randomness, irregularity, and periodicity of traffic flow at intersections, target flow prediction is related to various sequences. Through the introduction of the attention mechanism, highly related historical information may be connected for multi-component flow data in the final prediction. Compared to seasonal autoregressive integral moving average model, artificial neural network, and recurrent neural network, the experimental results demonstrated that the proposed MCA model can meet the accuracy and effectiveness of complex non-linear urban traffic flow prediction models.