Research Of Multi-Timestep Prediction Method Of Multi Highway Toll Station Exit Volume Based On Spatial Temporal Attention Mechanism

As the construction of highway road network approaches to saturation,the importance of using Intelligent Transportation System to improve highway operation and management,coping with the conflict between the rapidly increasing travel demand and the limited supply of road resource,has revealed obviously.Traffic volume is a fundamental parameter in ITS.Accurate prediction of traffic volume,especially for long term multi-timestep with short interval,remains to be a critical technical challenge.Highway network system,as a fully separated toll network system,has been constructed to achieve a high level of informatization.Large amount of historical traffic volume data,including recent data,have been collected for toll purpose,which provides significant support for data-driven models.However,the collection of recent data has a certain degree of delay.Therefore,it's necessary to add preprocessing step to impute the missing data to improve input data quality of the prediction modelDriven by massive data,inspired by forming mechanism of target volume,comprehensively considering multi-dimension and multi-variant correlation,as well as conditions restricted by reality,this paper conducts in-depth research on how to accurately predict exit volume of highway stations within a region in real time.The achievement of this paper can help provide fundamental support for improving the level of traffic control,management,and guidance.The research mainly includes the following parts1.To solve the missing problem of recent entrance volume data caused by transmission delay,an imputation model based on Generative Adversarial Network is proposed,with correlation between volume of missing point and volume of neighboring time and space as main consideration.Experiment result shows that the proposed model performs lower imputation error compared with common models,proving the proposed model's ability to help complement the dimension of input vector and ensure the quality of input data2.Using station exit data of the entire Guangdong Province,the geographical distribution of the source of exit volume of highway toll stations is analyzed.According to the average travel time of the arriving vehicles,toll stations are classified into 3 groups,which are short-range attraction,mid-range attraction and long-range attraction.By analyzing the distribution density of volume source of typical stations,apparent clustering phenomenon of source volume of single station,and strong correlation between source volume distribution and regional economic development level are found.The founding above provide inspiration for proposing a prediction model with entrance volume as main input and attention mechanism as key component.3.Analyzing correlation between exit volume and entrance volume from the aspect of micro traffic evolution and macro trend correlation,three main mechanism of partial spatial correlation,partial temporal correlation and integral spatial correlation are proposed.Time-varying characteristics is also found in the analysis.Furthermore,by analyzing hourly volume distribution of a station within a month,time dependent pattern and short-term continuity of toll station exit volume are also found.All results found above provide traffic knowledge preparation for the building of the prediction model.4.Based on the analysis result of affecting factors and the correlation mechanism,a multi toll station multi-timestep exit volume prediction model is built.The proposed model takes surrounding stations or potential highly correlated stations of target station as main input,and is equipped with the ability to intentionally learn strong correlated spatial and temporal features by adding attention mechanism to Encoder-Decoder framework embedded with LSTM network.External features including historical exit volume of target station and temporal attribute are also included as input.5.Critical region of Northern Guangzhou highway network,which consists of 14 target stations,is selected as experimental region.8 timesteps with 15-minute interval is to be predicted.By extracting and analyzing the final training result of spatial and temporal attention weight,the process of the model is properly explained,the interpretability of the model is improved.By comparing full-time error and step-wise error of different models,the effectiveness of the proposed framework and mechanism is proven.