Research On Time Series And Spatio-temporal Forecasting Of Traffic

As one of the negative impacts of urbanization,traffic congestion has many hazards and disadvantages,such as interfering with social operational efficiency and causing economic losses,causing environmental pollution and energy waste,and affecting residents’ travel experience and happiness.Therefore,people hope that the observed historical traffic data can be used to predict future trend changes,so as to better plan and make decisions about traffic travel.However,due to the influence of the city’s complex geographical environment,such as urban layout,road planning,population size and terrain conditions,it is a great challenge to achieve high-precision traffic flow prediction.Traditional traffic flow forecasting methods often fail to achieve good forecasting results when faced with complex and nonlinear multivariate forecasting tasks.Recently,data-driven deep learning models have attracted much attention due to their excellent performance in handling nonlinear tasks.The main research content of this paper is one-dimensional traffic time series prediction and two-dimensional traffic space-time prediction based on deep learning.Focusing on traffic time series and traffic grid spatio-temporal data sets,this study proposed two corresponding deep learning prediction models,and verified the effectiveness of the models by designing relevant comparative experiments.The main contributions of this paper are summarized as follows:1.For one-dimensional traffic time series prediction,this paper proposes a Multilevel Information Aggregation Network(MIANet)model,emphasizing the utilization,aggregation and balance of time information at different levels.The main innovations of MIANet include: 1)A new folded loop structure oriented to mixed periodic patterns,which is used to solve the utilization problem of mixed periodicity.2)A new type of recurrent unit called folded convolution aggregate temporal memory unit is used to solve the problem of multi-level information balance.3)The fusion decoder structure is used to solve the future information fusion problem and the prediction lag problem.4)A new skip-connection autoregressive linearstrategy to deal with the scale insensitivity problem existing in nonlinear networks.In the prediction comparison experiment of Traffic and Pedestrian two real open source datasets(a total of 24 indicators),MIANet’s prediction results achieved the best results on 16 indicators,and achieved sub-optimal results on 6 indicators.2.For two-dimensional traffic spatio-temporal prediction,this paper proposes a Multiscale Spatial Relation Aggregation Network(MSRANet)model,emphasizing the utilization,aggregation and balance of spatio-temporal relations at different scales.The main innovations of MSRANet include: 1)A new prediction framework that can effectively integrate hour-level adjacent time features,day-level cycle time features,and week-level cycle time features.2)Three novel spatial components are used to deal with physical spatial correlation(PCLSTM),functional spatial correlation(FCLSTM)and global functional spatial similarity(FSAM),respectively.3)A multi-task scaling strategy called exogenous variable regularization to better utilize exogenous variables.In the prediction comparison experiments of two real open source data sets,Taxi BJ and Bike NYC,the prediction results of MSRANet are better than those of the comparison prediction methods,and the evaluation errors have decreased by 4.8%(RMSE),2.0%(MAE),and 3.8%(RMSE)and 6.1%(MAE),respectively.