Short Term Traffic Flow Prediction Based On Deep Learning

Traffic flow data is usually nonlinear,highly random and time-dependent.Therefore,it is of practical significance to establish a short-term traffic flow prediction model based on the interaction of multiple factors.Based on convolution neural network and long-term and short-term time memory network,a short-term traffic flow prediction model is established by using deep learning method.The main research contents are as follows:(1)a TCN-LSTM model with causal convolution block is proposed,which is composed of two subnets: LSTM subnetwork is used to extract feature from original traffic flow data sequence,three TCN+LSTM subnetworks are used to extract features from traffic flow data with day-week-biweekly,holiday and weather.TCN is embedded to maintain causation of the input traffic flow data.Finally,features extracted from the two sub networks are merged and imported into top-level full connection network.The prediction sequence of the future short-term traffic flow is obtained at the output layer of FCN.(2)A new Conv-LSTM hybrid model is proposed.The partial convolution operation in CNN network is replaced by causal convolution to improve the extraction ability of temporal and spatial characteristics of traffic flow data.The structure of the model is divided into three parts: the first part designs a four layers convolution network,in which the convolution operation of the first and fourth layers is changed to causal convolution,which deals with the spatial characteristics of traffic flow data;the second part,uses LSTM＿1 extract Time feature;In the third part,the output of the two networks is spliced into LSTM＿2 layers output one-dimensional vector,and finally three layers output prediction results.The accuracy of the model is verified by the model performance evaluation standard and comparison with other models.Based on the traffic flow data set of Minneapolis St.Paul westbound road in Kaggle platform and the traffic network data set of California urban trunk road in PEMs database,MAE,MAPE,RMSE and coefficient of determination were used to evaluate the above two models.Compared with other models,the proposed TCN-LSTM and Conv-LSTM models have higher accuracy and stability in short-term traffic flow prediction.