Research On Traffic Flow Prediction Using Spatial-Temporal Fusion Methods

At present,as urban transportation becomes more convenient,the structure of urban road network is also becoming complex.Due to some unreasonable road planning or inadequate public transportation facilities,various traffic problems such as traffic congestion and traffic accidents have become increasingly apparent.Intelligent transportation system is an important means to solve the above problems,which can realize the optimal scheduling of road network,improve the efficiency and safety of traffic operation,and thus alleviate a variety of traffic problems.The traffic flow prediction problem is a key issue in the intelligent transportation system,which is related to the safety and stability of the transportation system.Initially,scholars often used statistical methods to solve the traffic flow prediction problem,which faced difficulties in data collection and large test errors.Subsequent developments of sensor technology has made it easier to collect traffic flow data,and the advancement of this technology has promoted the development of traffic flow prediction.In addition,the development of machine learning has also made traffic flow prediction tasks more efficient.Currently,the field of traffic flow prediction is further developing,and many researchers choose to use deep learning methods in this field.However,most of the methods proposed at present are focused on traffic flow prediction within one hour,neglecting the need for traffic flow prediction beyond one hour.To meet this requirement,a novel long-term time series prediction model is presented in this paper.In this paper,experiments were designed on two publicly available datasets and both achieved good results.The experiments demonstrate the effectiveness of the model.The main work of this paper is as follows:(1)To address the long-term traffic flow prediction beyond one hour,this paper proposes a long-term traffic flow prediction model based on Hybrid Spatial-Temporal Gated Convolution(HSTGCNN).In this paper,two main components,spatial-temporal attention mechanism and gated convolution,are designed to extract traffic flow features.(2)In this paper,a feature extraction idea of extracting global features first followed by longer-term features is designed.Dilated causal convolution is used in the gated convolution module,which effectively improves the long-term prediction capability of the model.(3)The experiments verified the effect of the HSTGCNN.HSTGCNN predicts the traffic conditions for 1 hour,1.5 hours,and 2 hours on two common traffic flow speed datasets,PeMS-bay and Los-loop.The experimental results show that HSTGCNN’s prediction accuracy is generally better than the compared baseline models such as ASTGCN and Graph WaveNet.In addition,this paper also conducted ablation experiments on the above two datasets and analyzed the effects of different modules in HSTGCNN.