Research On Traffic Speed Prediction Algorithm Based On Convolutional Neuralnetworks

Traffic prediction is an important issue in the study of intelligent transportation systems.Traffic speed is an important indicator to measure traffic status.Traffic speed prediction can not only provide scientific basis for traffic managers,but also provide support for other road services such as path planning.The traditional speed prediction models are mainly divided into two categories.One is a model that considers the traffic speed of a single road segment or a single detection station,such as LSTM,GRU,etc.,but the training time complexity of such model is high.And another is a model that considers the temporal and spatial correlations of traffic speed,but often ignores the influence of external factors such as weather,and often cannot obtain more accurate prediction results.In order to achieve more accurate traffic speed prediction,this paper proposes traffic speed prediction models based on temporal convolutional network.The main works of this paper are summarized as follows:(1)This paper proposes a traffic speed prediction model named Back-Propagation Temporal Convolutional Network(BTCN)for a single road segment or a single detector.This model can capture long-term temporal dependencies by using dilated causal convolution.And it is the first time that temporal convolution network is applied to traffic speed prediction.The experimental results show that the temporal convolutional network obtains more prediction accuracy than the traditional LSTM network,GRU network and the training speed of the BTCN model is 19.8% higher than that of the GRU network.(2)In order to capture the spatial dependences of traffic speed,this paper proposes a spatial-temporal prediction model named Temporal Graph Convolutional Network(TGCN),which consists of two components including spatial-temporal component and external component.Spatial-temporal component integrates graph convolutional network and temporal convolutional network.Firstly,the K-order graph convolutional network is used to capture spatially related features,and then these features are fed into a temporal convolutional network(TCN)to learn the characteristics of traffic speed over time.The external component considers the impact of various influencing factors on road traffic speed,including traffic time information(morning and evening peak hours,holidays)and weather.In addition,when constructing the sensor graph structure,this paper uses the point of interest(POI)around the sensor station and the road structure features to calculate the similarity of the sensors,and then adds it to the weight calculation process of the adjacency matrix of the sensor network graph.The proposed model TGCN is tested on two real datasets,and the feasibility of the model is verified.Several parameters affecting the performance of the model are discussed experimentally.In addition,the experimental results show that the RMSE of the TGCN model is reduced by 15% compared with the state-of-art prediction model STGCN,and the model's training time is reduced by 23.5% compared with the STGCN model.(3)Due to the electronic faults and regular maintenance of the traffic detection sensors,there may be missing values in the time series data collected by the sensors.This paper uses the method of tensor decomposition to complement the missing values in the dataset.The experimental results show that the prediction using the tensor decomposition method can increase the accuracy of the prediction.