Research On Short-Time Traffic Speed Prediction Model Based On Deep Learning

Traffic prediction plays an important role in solving urban traffic problems.Real-time and accurate traffic prediction is beneficial to planning routes and improving travel efficiency.In recent years,the theory of deep learning has achieved fruitful results.Because of its flexible model construction,high-dimensional data processing and generalization ability,many researchers apply it to the field of transportation.Deep learning enables researchers to process large data with high performance,so they have more time to focus on how to analyze the properties of the data itself and build models.Therefore,this paper uses deep learning to study the prediction of short-term traffic speed,which can be summarized as follows:(1)The temporal and spatial characteristics of the data in this paper are analyzed.In terms of time characteristics,it mainly explores the temporal correlation of traffic speed,and calculates the correlation coefficient by combining distance measurement and correlation coefficient measurement.It is found that there is a strong correlation between the traffic speed at the base time and the traffic speed near the base time,and the farther the time is,the lower the correlation is.In terms of spatial characteristics,the paper mainly explores the spatial correlation of traffic speed.Moran index is used to analyze the spatial autocorrelation of road traffic speed.It is found that the spatial autocorrelation of road traffic speed decreases with the increase of the distance between roads.(2)Two different short-term traffic speed prediction models are constructed based on whether external factors are considered.the first is the traffic speed prediction model of dynamic spatiotemporal graph convolution neural network DSCNN,that is,the road network is abstracted into graph structure,and the dynamic adjacency matrix is constructed by using graph structure itself and historical velocity data.the traffic speed at short time is predicted by combining graph convolution with short-term and short-term memory network and attention mechanism.Dynamic graph convolution neural network traffic speed prediction model is intended to dig deep traffic speed data spatiotemporal characteristics.Considering that the vehicle speed will be affected by the road structure,the area of the road section,the weather condition and the time of people’s travel,the second model is improved on the basis of DSCNN.On the one hand,the adjacency matrix is constructed with the point of interest to replace the previous simple matrix;on the other hand,the multi-source factor module is added to the model,which makes the factors considered in predicting the speed more comprehensive.The two models show different advantages on Qingdao taxi GPS trajectory dataset.The prediction accuracy of the model considering multi-source data is higher than that of the DSCNN model,the RMSE、MAE、MAPE index is 27.85%,35.92%,20.55% less than the DSCNN model,but the latter is more complex than the former model.So the training time cost is higher.Therefore,the first model is more suitable under the condition of insufficient exogenous data or saving labor requirements.Although the second model is more complicated for data processing and more complex,it ensures the accuracy of the final prediction.