Regional traffic flow combination prediction model considering virtual space of the road network

Accurate traffic flow forecasting is an important technical measure to alleviate traffic congestion. Since traffic flow has spatial and temporal characteristics, thus the adequate extraction of its spatio-temporal features is an important prerequisite to promote the forecast accuracy of the model. However, a majority of existing traffic flow prediction models cannot sufficiently consider the neighborhood spatio-temporal relationship for the real road network in the modeling process, which makes it difficult to improve the model prediction accuracy. For this reason, this paper takes improved feature enhancement graph convolution (FEGC), gated recurrent unit (GRU), and improved lightweight particle swarm optimization (ILPSO) algorithm as components, respectively, to construct a combinatorial traffic flow prediction model FEGC-GRU-ILPSO (FGI), aiming to achieve accurate forecast for regional traffic flow through fully learning spatio-temporal correlation characteristics. Firstly, considering that most traffic flow modeling methods are ineffective in characterizing the hidden association information within nodes, we propose the method of constructing the virtual space adjacency matrix based on the improved gray relational analysis (IGRA) algorithm, which achieves the effective characterization of road network neighborhood relationship by fusing it with the original adjacency matrix. Then, based on the idea of matrix decomposition, the weight adjacency matrix is further introduced to realize the dynamic capture of time-varying correlation of node graph structure in realistic road networks. Secondly, to address the performance degradation problem caused by feature assimilation in multi-layer graph convolution, an improved feature enhancement graph convolution component is proposed to alleviate the multi-layer graph convolution over-smoothing by enhancing salient features. Finally, considering the convex optimization problem caused by the way the hyperparameters of the model are determined through subjective experience, we propose the ILPSO algorithm to improve the overall prediction performance in an adaptive optimizing method. In this paper, real-world data acquired by the Caltrans Performance Measurement System (PeMS) is used as the object of study. The experimental results demonstrate that the FGI model has better prediction performance than the current mainstream baseline models.