Short-Term Traffic Speed Prediction Based on AGC-LSTM with Multi-Source Data Integration

• Published in: International journal of ITS research 2024-12, Vol.22 (3), p.774-784
• Main Authors: Chen, Yujia, Gao, Mingxia, Xiang, Wanli, Mo, Junwen
• Format: Article
• Language: English
• Subjects: Accuracy; Automotive Engineering; Civil Engineering; Computer Imaging; Data integration; Electrical Engineering; Engineering; Pattern Recognition and Graphics; Prediction models; Road conditions; Robotics and Automation; Root-mean-square errors; Spatiotemporal data; Speed limits; Traffic speed; Urban environments; User Interfaces and Human Computer Interaction; Vision; Weather
• ISSN: 1348-8503; 1868-8659
• DOI: 10.1007/s13177-024-00431-2

The state changes of short-term traffic speeds in rapidly changing urban environments depend not only on historical data and their own patterns, but also closely relate to other variables. Although existing studies have incorporated weather factors and surrounding points of interest (POIs) into prediction models, they are still insufficient to fully capture the complexity and dynamic changes of traffic speed. In this paper, we propose a graph convolutional LSTM-based urban short-duration traffic speed prediction model (EF-AGC-LSTM) under multi-source data fusion that takes into account the road condition information on the basis of the existing weather and POI factors, embeds the GCN into the gating computation of the LSTM so as to acquire the spatio-temporal information simultaneously at each time step, and at the end introduces an attentional mechanism to identify and enhance the external key features. First, the model carefully measures things like weather, points of interest (POI), road conditions, and more. Next, it makes it possible for all of these complicated pieces of information to be combined in a useful way using a specially designed external feature set fusion component (EF-component). We then use the graph convolutional LSTM network (AGC-LSTM), which combines the attention mechanism, to extract spatio-temporal features for more accurate traffic speed prediction. Experimental comparisons with traditional models, considering only weather and POI, on real datasets show that the EF-AGC-LSTM model achieves better prediction performance with the introduction of road condition factors, and its mean absolute error (MAE), root mean square error (RMSE), and accuracy are improved. The EF-AGC-LSTM model outperforms the other comparison models in terms of MAE, RMSE, and accuracy. It indicates that EF-AGC-LSTM has excellent performance in capturing dynamic traffic speed changes.