Research On Spatio-Temporal Prediction Method Of Traffic Flow Based On Deep Learning

Urban traffic congestion has always been one of the main contradiction that hinders social and economic development,and traffic flow prediction is a commonly used technique to solve this contradiction.In view of the complex nonlinearity of the transportation system and the characteristics of being easily disturbed by uncertain factors,the establishment of a prediction model that can cope with the influence of uncertain factors and effectively evolve the nonlinear spatiotemporal variation law from a macro and micro perspective is of great significance to assist the traffic management department in alleviating traffic congestion,optimizing management and control services and responding to emergencies.From the perspective of the research on the external spatiotemporal characteristics and internal chaotic characteristics,this thesis proposes a deep learning-based spatiotemporal prediction method system and research framework for traffic flow,and takes “spatiotemporal fusion data repair-longitudinal and horizontal correlation spatiotemporal prediction under time lag neighborhood correlation spatiotemporal prediction under the influence of weather-chaos analysis and prediction of measured traffic flow data” as the main line to carry out relevant research.Details are as follows:(1)Aiming at the problem of coarse-grained feature mining in the spatiotemporal correlation repair method,a spatiotemporal graph convolution traffic data repair method is proposed,combining graph convolution,one-dimensional convolution and correlation theory,from the model input data organization level,the spatiotemporal fusion matrix is constructed through fusion rules to realize the fine-grained description of graph convolutional adjacency relationship,and it is aggregated with the spatiotemporal matrix information of multi-channel historical traffic flow to realize the fine-grained spatiotemporal feature mining at the model processing level.In the random missing data mode of the road network,this method can effectively extract the rich spatiotemporal features of historical series,so as to complete the task of missing data repair in the preprocessing stage,and provide a high-quality data basis for subsequent predictive modeling.(2)Aiming at the problems of coarse-grained feature mining and insufficient consideration of spatiotemporal characteristics caused by the unrefinement of the spatial arrangement relationship of the road network,this thesis proposes a spatiotemporal prediction method for the longitudinal and horizontal correlation under time lag from the perspective of the spatial time lag of external spatiotemporal characteristics,the longitudinal correlation of upstream and downstream sections,and the horizontal correlation of multi-lanes.In terms of longitudinal spatiotemporal relationship,the bidirectional long-term and short-term memory network is introduced.On the basis of quantifying and eliminating the influence of time lag,a data input organization method of vector splitting is proposed,and a spatiotemporal liquidity characteristic expression of bidirectional dimension is established.In terms of the horizontal spatiotemporal relationship,the temporal convolution network is introduced,and the dilated causal convolution is used to realize the collaborative extraction of multi-lane horizontal spatiotemporal features to solve the defect of splitting and extracting spatiotemporal features in traditional modeling.Finally,the attention mechanism is used to dynamically allocate the contribution of two types of spatiotemporal features to the output.This method integrates the influence of three types of spatiotemporal characteristics,can relatively complete the evolution of complex time lag longitudinal and horizontal spatiotemporal relationships.(3)Aiming at the problem that existing methods are difficult to effectively describe the nonEuropean spatial heterogeneity of the real road network,and the performance of the model is degraded by uncertain weather disturbance.This thesis proposes a neighborhood-dependent spatiotemporal prediction method under the influence of weather from the perspective of the neighborhood correlation of external spatiotemporal characteristics.This method defines a new adjacency relationship to optimize the embedding of graph structure,solves the problem of actual node space deformation caused by over dependence on parameter training,and proposes a hierarchical residual structure to solve the problem of node feature convergence caused by network deepening and over smoothing.With the advantage of the graph convolution-gated recurrent unit,a sequence-to-sequence neighborhood correlation multi-step prediction framework is established,and a weather flow decision module is designed to enhance the anti-interference ability of the model.This method realizes the effective extraction of non-European spatiotemporal relationship of the real road network,and has obvious advantages in the long-term prediction of multi-step prediction,and the model still has stable performance under weather disturbance.(4)In order to solve the problem that the saturation correlation dimension method cannot take into account the accuracy and efficiency when determining the parameters,and the chaotic modeling is not aimed at phase space data structure and gradient descent parameter update method is easy to fall into local optimum.From the perspective of internal chaos research,this thesis proposes to improve the saturated correlation dimension method of norm fusion search domain to determine the embedding dimension,combine the delay time obtained by mutual information method to complete high-quality phase space reconstruction,and judge the chaos of measured traffic flow series by small amount of data method.Then,a lightweight improved particle swarm optimization algorithm is proposed,and a chaos prediction model combining deep belief network and gated recurrent unit is established based on the phase space data structure,and the improved particle swarm optimization algorithm is used to replace the traditional gradient descent parameter update method to realize the accurate prediction of chaotic traffic flow,so as to analyze the internal nonlinear dynamic operation mechanism of the measured traffic flow data from a micro perspective.This research provides a systematic research idea for deep learning traffic spatiotemporal prediction modeling,enriches the research content of artificial intelligence technology deep integration of intelligent traffic prediction,and provides decision-making basis and method support for traffic congestion control and emergency response to sudden weather.