Short-term traffic flow prediction: From the perspective of traffic flow decomposition

Abstract

Some researchers treat traffic flow as an entirety while predicting short-term traffic flow. Through analyzing real-world traffic flow, we have found that urban traffic shows a stable changing process along with random disturbs. An alternative way is to decompose traffic flow into two components: periodicity and volatility. We propose a hybrid method named Time-Series Analysis and Supervised-Learning (TSA-SL) for short-term traffic flow prediction from the perspective of traffic flow decomposition. In the method, period traffic flow is modeled with a typical TSA method called Fourier Transform (FT), where periodic behaviors are described as the combination of sines and cosines. The volatility of the current location is determined by its surroundings, so spatial–temporal correlations are extracted as input features of SL. Then, three hybrid prediction models, including FT-SVR, FT-GBRT and FT-LSTM, are built with proposed TSA-SL. In the experiment, an Electronic Registration Identification (ERI) dataset including massive real-world individual trajectories is employed. Comparing with classical baseline models, our proposed TSA-SL method has certain superiority. Furthermore, we decompose traffic flow into different components in terms of traveling purposes and vehicle types. The experimental results show that our method performs better in predicting partial traffic flow than predicting all traffic flow.