A Study Of Traffic Forecasting Based On Deep Spatiotemporal Models Over Graph

With the acceleration of building China’s strength in transportation,vast traffic management methods and decision-making methods are shifting from traditional factor driven to innovative technology driven.As a basic problem in traffic management,the solution of traffic flow prediction also has obvious changes,i.e.,it has gradually developed from the original single time series model construction to the deep spatiotemporal model design by combining spatial and temporal characteristics.Although the existing models have made considerable progress compared with the classical methods,there are still some shortcomings in the solutions of traffic flow prediction.In areas with abundant data resources,extracting and fusing multiple data features play an important role.In areas with insufficient data resources,it is meaningful to design deep spatiotemporal models to capture complex spatiotemporal features from few samples and quantify uncertainty.First of all,in order to solve the challenge of capturing and fusing multiple data features for regions with rich data resources,especially in terms of extracting static spatial dependence and external features,this article constructs a multi-feature spatiotemporal convolutional network model.The proposed model includes the spatiotemporal feature extraction module and the external feature extraction module to extract complex spatiotemporal features and external features separately,also uses the strategy of dynamically assigning feature weights to fuse multiple data features.Finally,the fully connected layers of the proposed model predict the traffic flow.This article creatively proposes that the Bayesian optimization is used to quantify static spatial dependence,combined it with the spatiotemporal feature extraction module to extract complex spatiotemporal features.External features are quantified external features by embedded technology,thereby solving the shortcomings of existing deep spatiotemporal models that only acquire spatiotemporal features.In addition,the attention mechanism is used to dynamically assign feature weights,which can reflect the dynamic impact of different features on traffic flow.Thus,in a real scene with multifeatures,such as traffic flow,vehicle type,and data type,the proposed model can accurately predict traffic flow on multiple monitoring points or on a single monitoring point,and traffic prediction during holidays.Secondly,for areas with scarce data resources,in order to extract complex spatiotemporal features from a small amount of data to accurately predict traffic flow and quantify spatiotemporal uncertainty.This article proposes a deep graph gaussian process method.The proposed method stacks the aggregation gaussian process,the temporal convolutional gaussian process,and the gaussian process with linear kernel function,and infers the model parameters based on the improved inference algorithm.In the proposed method,the creatively proposed aggregation gaussian process to solve the problem that existing gaussian processes or deep gaussian processes cannot accurately obtain the dynamic spatial dependence.At the same time,by stacking the aggregation gaussian process and the temporal convolutional gaussian process that can obtain the temporal features,it is used to extract complex spatiotemporal features.And then the extracted features are given by the gaussian process with linear kernel function to predict the traffic flow.The deep spatiotemporal model built based on the deep graph gaussian processes not only has a simple design and implementation process,but also provides accurate traffic flow predictions with only a small amount of data on multiple real datasets.In addition,the constructed model also provides uncertainty measures,thereby reducing the impact of uncertainty on future decisions.