Research On Road Traffic Prediction Application Based On Spatiotemporal Attentio

With the rapid development of the socio-economy,road traffic conditions become more and more complex,which leads to traffic congestion and other problems becoming particularly prominent.Traffic flow prediction,as one of the key research directions in intelligent traffic management and decision-making,can help people to intervene in advance to solve road problems such as traffic congestion.So far,many relevant prediction methods have been applied to traffic flow prediction,but the inherent complexity,road network correlation,spatiotemporal nonlinearity,and human participation uncertainty in traffic systems overlap to make traffic flow prediction a challenging task.Therefore,to effectively model the latent correlations and their trends in spatiotemporal traffic data,deep learning approaches based on spatial-temporal attention mechanism are proposed to predict traffic flows.The main work of the paper is described as follows.To address the problem that a single model cannot effectively capture the complex spatiotemporal nonlinearities of traffic data,which makes it difficult to obtain satisfactory prediction results.We propose an Attention-based Gated Convolutional-Bidirectional Long Short-Term Memory(AGC-BLSTM)network model for traffic flow prediction.The AGC-BLSTM model can capture the spatial and temporal features of traffic flow through a multi-layer network structure consisting of Gated Convolutional Neural Network and Bidirectional Long Short-Term Memory network.The attention mechanism can automatically identify the most important historical moments for the current prediction to assign corresponding weights to the traffic features extracted at different moments.In addition,the model exploits parallel sub-modules to model three temporal attributes of traffic flow,i.e.,weekly,daily,and short-term time dependencies,and finally fuses the outputs of these three modules to obtain the prediction results.The experiment results on the road traffic dataset of Guiyang city show that the AGC-BLSTM model has better prediction results compared with other benchmark models.Considering the high dynamic spatiotemporal correlation and road network correlation of traffic flow data,such as the degree of interaction between different locations at any moment is different,and the correlation between different locations at the same moment is various,we propose a Spatial-Temporal Attention-based Road Traffic Prediction(STARTP)model.In the STARTP model,the dynamic connections of different locations in spatial dimensions are adaptively captured by the encoder module with spatial attention,modeling the spatial correlation and period dependence simultaneously.The temporal dependences at different moments are captured by the decoder module with temporal attention.Finally,the linear relationship of the traffic flow is extracted by combining the Autoregressive model to obtain the prediction results.The experiments on Pe MSD4 and Pe MSD8 datasets verify that the STARTP model can effectively capture the dynamic spatiotemporal features of road traffic data with better prediction results.In summary,the AGC-BLSTM and STARTP models proposed in the paper,based on attention mechanisms,can better extract spatiotemporal fusion features and explore dynamic spatiotemporal correlations.Compared to traditional time series analysis models and advanced deep learning models,they have lower prediction errors,which is of great significance for improving the efficiency of intelligent transportation management.