PCR: A Parallel Convolution Residual Network for Traffic Flow Prediction

Traffic flow prediction is crucial in smart cities and traffic management, yet it presents challenges due to intricate spatial-temporal dependencies and external factors. Most existing research relied on a traditional data selection approach to represent temporal dependence. However, only considering spatial dependence in adjacent or distant regions limits the performance. In this paper, we propose an end-to-end Parallel Convolution Residual network (PCR) for grid-based traffic flow prediction. First, we introduce a novel data selection strategy to capture more temporal dependence, and then we implement an early fusion strategy without any additional operations to obtain a lighter model. Second, we propose to extract external features with feature embedding matrix operations, which can represent the interrelationships between different kinds of external data. Finally, we build a parallel residual network with concatenated features as input, which is composed of a standard residual net (SRN) to extract short spatial dependence and a dilated residual net (DRN) to extract long spatial dependence. Experiments on three traffic flow datasets TaxiBJ, BikeNYC, and TaxiCQ exhibit that the proposed method outperforms the state-of-the-art models with the most minor parameters.