Short-Term Traffic Flow Prediction with Weather Conditions: Based on Deep Learning Algorithms and Data Fusion

• Published in: Complexity (New York, N.Y.) Vol. 2021; no. 1
• Main Authors: Hou, Yue, Deng, Zhiyuan, Cui, Hanke
• Format: Journal Article
• Language: English
• Published: Hoboken Hindawi 2021; John Wiley & Sons, Inc; Wiley
• Subjects: Accuracy; Algorithms; Data integration; Deep learning; Intelligent transportation systems; Machine learning; Meteorological data; Methods; Neural networks; Optimization; Prediction models; Principal components analysis; Radial basis function; Support vector machines; Time series; Traffic congestion; Traffic flow; Traffic information; Traffic models; Weather
• ISSN: 1076-2787; 1099-0526; 1099-0526
• DOI: 10.1155/2021/6662959

Short-term traffic flow prediction is an effective means for intelligent transportation system (ITS) to mitigate traffic congestion. However, traffic flow data with temporal features and periodic characteristics are vulnerable to weather effects, making short-term traffic flow prediction a challenging issue. However, the existing models do not consider the influence of weather changes on traffic flow, leading to poor performance under some extreme conditions. In view of the rich features of traffic data and the characteristic of being vulnerable to external weather conditions, the prediction model based on traffic data has certain limitations, so it is necessary to conduct research studies on traffic flow prediction driven by both the traffic data and weather data. This paper proposes a combined framework of stacked autoencoder (SAE) and radial basis function (RBF) neural network to predict traffic flow, which can effectively capture the temporal correlation and periodicity of traffic flow data and disturbance of weather factors. Firstly, SAE is used to process the traffic flow data in multiple time slices to acquire a preliminary prediction. Then, RBF is used to capture the relation between weather disturbance and periodicity of traffic flow so as to gain another prediction. Finally, another RBF is used for the fusion of the above two predictions on decision level, obtaining a reconstructed prediction with higher accuracy. The effectiveness and robustness of the proposed model are verified by experiments.