Urban Traffic Flow Prediction Using a Gradient-Boosted Method Considering Dynamic Spatio-Temporal Correlations

Abstract

Traffic flow prediction at the city scale is an important topic useful to many transportation operations and urban applications. However, this is a very challenging problem, affected by multiple complex factors, such as the spatial correlation between different locations and temporal correlation among different time intervals. Considering spatio-temporal correlations, we propose a novel short-term traffic flow prediction model, which combining a gradient–boosted regression tree model and principal component analysis, and we name our model as DSTGBRT. First, we analyze the spatio-temporal correlations among different locations, using Pearson correlation coefficient. Second, we combine Pearson correlation coefficient and historical traffic flow data to construct feature vector and get original training data. Third, to eliminate the linear correlation between features, we use principal component analysis to construct new feature vector and get new training data. In the experiments, compared with traditional spatio-temporal gradient–boosted regression tree model named as STGBRT, the results demonstrate that our proposed DSTGBRT can do a timely and adaptive prediction even in the rush hour when the traffic conditions change rapidly. At the same time, it outperforms the existing methods.