Research And Implementation Of A Multi-Step Traffic Index Prediction Tool Oriented To Time-Varying Topological Relations

The construction of complex traffic road networks makes traffic accidents prone and brings a lot of inconvenience to people’s travel,thus it is crucial for traffic index forecasting.Traffic index prediction is divided into single-step prediction and multi-step prediction,which refers to predicting traffic index at a certain moment and multiple moments in the future,respectively.The challenges of traffic index forecasting are reflected in the following aspects:(1)the topological relationship of the traffic network changes with road maintenance or congestion,and this dynamic relationship is also called time-varying topological relationship,which is difficult to be captured accurately;(2)in multi-step forecasting,the prediction error becomes larger as the number of prediction steps increases,which is also called the error propagation problem;(3)due to the small density of road sensors and sensor failure,traffic data will become sparse,affecting the model prediction performance,which is also known as the data sparsity problem.To address the above challenges,this paper designs and implements a multi-step traffic index prediction tool for time-varying topological relations,and the main contributions of this paper are as follows:(1)For time-varying topological relations that are difficult to capture,this paper proposes a time-varying adjacency mask.It includes two parts:self-adaptive semantic adjacency matrix and mask.The self-adaptive semantic adjacency matrix is fixed,which is not in line with the dynamic nature of the topological relations,so the mask is proposed to correct it,so that the topological relations obtained at each moment are different but the phase difference is not very high.(2)To address the error propagation problem arising from multi-step prediction,self-smoothing regularization is proposed in this paper.It can use prior knowledge to learn the difference values between adjacent prediction steps,so as to limit the difference values predicted by the model and play the role of regularization.(3)For the problem of sparse data,this paper proposes a spatiotemporal decay expansion strategy.It uses the connectivity between roads for spatial expansion to complete the missing data,and uses the historical data of roads for temporal expansion to complete the missing data,and also sets the spatio-temporal decay coefficient to indicate the confidence of the completed data.(4)Based on the above contributions,a multi-step traffic speed prediction algorithm and a single-step traffic condition prediction algorithm are implemented in this paper.Based on the above algorithms,a fine-grained and high-precision traffic index multi-step prediction tool is designed and implemented in this paper,which can provide historical data display,single-step prediction and multi-step prediction of traffic index.