Short-term traffic prediction based on time series decomposition

Abstract

Traffic flow decomposition is an alternative method to explore the composition of traffic flow and improve prediction accuracy. However, most of them suffer from the inability to fully utilize the character of traffic data. This paper presents a novel framework for traffic flow decomposition and modeling named Time Series Decomposition (TSD). The traffic flow is adaptively decomposed into periodic component, residual component and volatile component which are modeled respectively. Empirical Mode Decomposition (EMD) is applied to extract the intrinsic mode functions (IMFs) of traffic flow, the periodic patterns are intuitively presented via Hilbert transform in terms of frequencies. Then the periodic component can be described as a Fourier series based on obtained frequencies. Meanwhile, the residual component is presented by IMF with the lowest frequency. The remaining part is the volatile component modeled by supervised learning. The proposed hybrid model is evaluated on the real-world dataset and compared with classical baseline models. The results demonstrate that TSD can unearth the underlying periodic patterns and provide an explicable composition of the traffic flow. Furthermore, the volatile component ensures the accuracy of single-step prediction while periodic and residual components show promising abilities in improving the multi-step prediction accuracy of short-term traffic flow.