Research on Traffic Flow Prediction Based on Spatio-Temporal Correlation Analysis

With the advancement of urbanization, people’s living standards have significantly improved, but have also resulted in various challenges, particularly in the domain of transportation. Traffic congestion and low traffic efficiency have led to substantial wastage of people’s travel time. Traffic flow prediction is vital to Intelligent Transportation Systems (ITS) as it centers on projecting the future condition of road networks using historical data and pertinent information. Accurate traffic flow prediction helps users take proactive measures and is essential for addressing traffic challenges. Yet, the intricate nonlinear temporal and spatial correlations within traffic flow data pose notable challenges. To address the intricate correlations within traffic flow data, this study separately investigates the temporal and spatio-temporal correlations. Firstly, a model based on time correlation is developed, followed by the construction of a new network model that integrates time and space to extract complex spatio-temporal correlations from traffic flow data. The key research focuses are as follows: the analysis of time correlation in traffic data, the introduction of several commonly used neural network models for time sequences, and the incorporation of gating mechanisms into the TCN to create an enhanced gating TCN network structure model. This model aims to analyze the correlation between traffic data and time. The proposed gating mechanism TCN involves significant changes in network structure, with the integration of input and output gates akin to the LSTM structure. Furthermore, the convolutional correlation modules in each extended convolutional module are substituted by two internal parallel convolutional modules in the TCN. Forming an input gate and output gate structure. Furthermore, in order to reduce variance and simultaneously increase the input and output gates, Each parallel convolution component is enhanced by the addition of two identical parallel branches, resulting in the total output being the average of all the “gate” outputs. The performance of the proposed model is assessed using real traffic speed datasets. Demonstrating superior prediction accuracy and effectively capturing sudden changes in traffic speed while maintaining stable results.