Wanfeng Chang

Due to the regular influence of human activities, traffic flow data usually exhibit significant periodicity, which provides a foundation for further research on traffic flow data. However, the temporal dependencies in traffic flow data are often obscured by entangled temporal regularities, making it challenging for general models to capture the intrinsic functional relationships within the data accurately. In recent years, a plethora of methods based on statistics, machine learning, and deep learning have been proposed to tackle these problems of traffic flow forecasting. In this paper, the Transformer is improved from two aspects: (1) an Efficient Attention mechanism is proposed, which reduces the time and memory complexity of the Scaled Dot Product Attention; (2) a Generative Decoding mechanism instead of a Dynamic Decoding operation, which accelerates the inference speed of the model. The model is named EGFormer in this paper. Through a lot of experiments and comparative analysis, the authors found that the EGFormer has better ability in the traffic flow forecasting task. The new model has higher prediction accuracy and shorter running time compared with the traditional model.

Accurate traffic flow forecasting is very important for urban planning and traffic management. However, this is a huge challenge due to the complex spatial–temporal relationships. Although the existing methods have researched spatial–temporal relationships, they neglect the long periodic aspects of traffic flow data, and thus cannot attain a satisfactory result. In this paper, we propose a novel model Attention-Based Spatial–Temporal Convolution Gated Recurrent Unit (ASTCG) to solve the traffic flow forecasting problem. ASTCG has two core components: the multi-input module and the STA-ConvGru module. Based on the cyclical nature of traffic flow data, the data input to the multi-input module are divided into three parts, near-neighbor data, daily-periodic data, and weekly-periodic data, thus enabling the model to better capture the time dependence. The STA-ConvGru module, formed by CNN, GRU, and attention mechanism, can capture both temporal and spatial dependencies of traffic flow. We evaluate our proposed model using real-world datasets and experiments show that the ASTCG model outperforms the state-of-the-art model.