Zhu Rong, Tang Jie, He Xuansen, Zhou Xianlai, Huang Xiaohui, Wu Fengyun, Chen Songli
School of Modern Information Industry, Guangzhou College of Commerce, Guangzhou, China.
Guangdong Embedded Information Technology Industry College, Guangzhou Xinhua University, Guangzhou, China.
PLoS One. 2024 Dec 5;19(12):e0312601. doi: 10.1371/journal.pone.0312601. eCollection 2024.
Traffic flow prediction is an important part of transportation management and planning. For example, accurate demand prediction of taxis and online car-hailing can reduce the waste of resources caused by empty cars. The prediction of public bicycle flow can be more reasonable to plan the release and deployment of public bicycles. There are three difficulties in traffic flow prediction to achieve higher accuracy. Firstly, more accurately to capture the spatio-temporal correlation existing in historical flow data. Secondly, the weight of each channel in the traffic flow data at the same time interval affects the prediction results. Thirdly, the proportion of closeness, period and trend of traffic flow data affects the prediction results. In this paper, we design a deep learning algorithm for short-term traffic flow prediction, called SE-MAConvLSTM. First, we designed Spatio-Temporal Feature Extraction Module (STFEM), which is composed of Convolutional Neural Network (CNN), Squeeze-and-Excitation Network (SENet), Residual Network (ResNet) and Convolutional LSTM Network (ConvLSTM) to solve the above two problems mentioned. In addition, we design multi-attention modules (MAM) to model the closeness, period and trend of traffic flow data to solve the third problem mentioned above. Finally, the aggregation module was used to integrate the output of the last time interval in STFEM and the output of the multi-attention module. Experiments are carried out on two real data sets, and the results show that the proposed model reduces RMSE by 4.5% and 3.7% respectively compared with the best baseline model.
交通流预测是交通管理与规划的重要组成部分。例如，准确预测出租车和网约车的需求可以减少空驶造成的资源浪费。对公共自行车流量的预测能够更合理地规划公共自行车的投放与部署。交通流预测要实现更高的准确率存在三个难点。首先，更准确地捕捉历史流量数据中存在的时空相关性。其次，同一时间间隔内交通流数据中各通道的权重会影响预测结果。第三，交通流数据的紧密性、周期性和趋势性所占比例会影响预测结果。在本文中，我们设计了一种用于短期交通流预测的深度学习算法，称为SE-MAConvLSTM。首先，我们设计了时空特征提取模块（STFEM），它由卷积神经网络（CNN）、挤压激励网络（SENet）、残差网络（ResNet）和卷积长短期记忆网络（ConvLSTM）组成，以解决上述提到的两个问题。此外，我们设计了多注意力模块（MAM）来对交通流数据的紧密性、周期性和趋势性进行建模，以解决上述第三个问题。最后，使用聚合模块对STFEM中最后一个时间间隔的输出和多注意力模块的输出进行整合。在两个真实数据集上进行了实验，结果表明，与最佳基线模型相比，所提出的模型分别将均方根误差（RMSE）降低了4.5%和3.7%。
