Research On Regional Traffic Prediction Based On Deep Learning Combination Modeling

Traffic prediction is one of the essential technologies in intelligent transportation system.Accurate and real-time traffic flow prediction can provide theoretical basis for road traffic management departments to implement control measures and provide reasonable travel suggestions for traffic participants,so as to optimize the use of road network resources.It is one of the main means to help solve the problem of urban traffic congestion.In view of the complex characteristics of the road network in reality and the sensitivity of the urban road traffic system to environmental changes,the traffic flow data is expressed as highly nonlinear time series data.In the current research,the data-driven method has become the main modeling method in the field of traffic flow prediction because of its high applicability to complex nonlinear problems,which mainly depends on the implementation of deep learning models.The development of data awareness technology and artificial intelligence technology has promoted the generation of traffic big data.How to mine effective information from large-scale and complex traffic flow data by deep learning,realize the future space-time situation awareness of the road network,and provide reliable services for traffic management departments and public travel is one of the major challenges in the field of intelligent transportation.Based on the characteristics of heterogeneous and multivariate spatio-temporal data of traffic flow,this paper realizes regional prediction modeling.Combined with deep learning combination modeling theory,optimization theory in graph convolution network and data fusion technology,a novel research framework of spatio-temporal situation awareness,analysis and application of regional road network driven by multivariate spatio-temporal data is proposed.Firstly,it combs the development context of traffic flow prediction,summarizes the problems and challenges in the current field of traffic flow prediction,clarifies the research focus and basic direction,and lays a basic idea for the research content of this paper.Secondly,it puts forward the data processing scheme of multivariate spatio-temporal traffic data,analyzes the characteristics of the current data set,analyzes the spatio-temporal characteristics and multivariate parameter disturbance characteristics of traffic flow data based on the processed data,grasps the main concerns of characteristic modeling,and provides data analysis basis for the establishment of prediction model.Then,based on the spatiotemporal characteristics of traffic flow,a graph convolution gated recurrent unit which is suitable for spatio-temporal data mining is constructed to realize the regional spatio-temporal prediction of traffic flow;The multi-head graph attention mechanism is introduced to optimize the quantitative expression of spatial correlation and build a more accurate spatio-temporal relationship of road network;Aiming at the over smoothing phenomenon in the application of deep graph convolution,a multi-scale graph convolution algorithm is proposed to optimize the structure in multi-layer model to improve the performance of graph convolution network.Finally,the accuracy and robustness of spatio-temporal prediction are further improved through the prediction with decision-level data fusion based on the multivariate data;The single parameter based prediction method can not effectively identify the nonlinearity and instability in the traffic flow data,therefore,the correlation prediction model is established independently according to the internal parameter correlation characteristics of the traffic flow and the external environmental disturbance,and the multiple support decision-making is constructed;Aiming at the problem that multivariate data is difficult to meet the needs of combined modeling with different parameter characteristic,a decision level significant linear fusion prediction algorithm is proposed,which reduces the complexity of multivariate data in deep learning combined modeling at the level of decision-making level fusion,improves the reliability,robustness and accuracy of prediction results through the fusion and reconstruction of predicted values,and provides a new modeling idea for building a intelligent decision-making system supported by multivariate data.The research provides a systematic research idea and technical basis for regional prediction based on multi-dimensional and heterogeneous traffic spatio-temporal data,constructs a regional traffic flow prediction model with high accuracy and strong universality,promotes the development of spatio-temporal situational awareness in intelligent transportation system,and can help alleviate the problem of urban traffic congestion.