Research On Short-term Traffic Prediction Algorithm Based On Deep Learning

As one of the basic problems in the mobility on demand(MoD)system,traffic prediction is the premise of the smooth operation of the intelligent transportation system and the solution to the problem of declining residents' quality of life and the urbanization of society caused by the rapid increasing.number of vehicles in the transportation network.In recent years,the data modeling algorithm based on deep learning has attracted the attention of many scholars,and its broad application provides a new way to solve the traffic prediction problem.However,deep.learning-based algorithms still have,the following limitations in the field of traffic prediction:Though long short-term memory(LSTM)recurrent neural networks can fully extract the dependence of sequence data on the time dimension,this method is difficult to model the internal spatial correlation of traffic data.In addition,because the data sampling of convolutional neural networks(CNNs)is fixed in the geometric region,although it can fully extract the spatial dependence of data,it still leads to insufficient feature extraction when modeling the spatial dependence of traffic data.In order to overcome the above-mentioned limitations,this paper extends the long short-term memory recurrent neural networks andconvolutional neural networks,and combines.deep learning technologies such as graph attention mechanism to solve the travel demand forecasting problem and traffic flow prediction problem.The main research contents and research results of the paper are summarized as follows:1.An improved long short-term memory recurrent neural network based travel demand forecasting model is proposed.To improve the ability to model the spatio-temporal dependence of travel demand data,the model uses a convolutional neural network to replace the matrix multiplication in the long short-term memory cell to achieve the simultaneous extraction of spatio-temporal correlation.To further improve forecasting performance,the model has been redefined for the travel demand forecasting problem by introducing residual connection.In addition,this paper proposes demand request vectors(DRV)based input data repre sentation to improve its feature representation capabilities.Compared with the baseline algorithms,the experimental results show that the model has good prediction performance,and can model the spatio-temporal dependence of the data and the influence of external factors simultaneously.2.A travel demand prediction model based on improved.temporal convolutional neural network and a feature enriching module achieved by a deformable convolutional neural network is proposed.Modeling the spatio-temporal dependence of the data is the premise of realizing travel demand forecasting.This model extends the temporal convolutional neural network to enable it to adapt to serial travel demand data with spatial correlation,and a travel demand forecasting model is built based on it.In order to further improve the feature representation ability of the model,this paper builds a feature enriching module based on a deformable convolutional neural network,breaking the limitations of the traditional convolutional neural network on the data sampling process and improving the model performance.Compared with the baseline algorithms,the experimental results show that the model has smaller prediction errors.Meanwhile,compared with the baseline algorithms based on recurrent neural networks,the training time of the model is shorter,which can save the training resources.3.A traffic flow prediction model based on long short-term memory recurrent neural network and graph attention mechanism is proposed.According to the spatial distribution characteristics of traffic flow data,we use an undirected graph to represent the traffic network and build a traffic flow prediction model based on it.To model the spatial-temporal dependence of traffic flow data,the model uses a graph attention mechanism to replace the matrix multiplication in the long short-term memory cell.Meanwhile,in order to further improve the model's ability to model the spatial dependencies,this model extends the graph attention mechanism to achieve the re-correction of the space dependencies.Compared with the baseline algorithms,the experimental results show that the model can realize real-time prediction of traffic flow with less error.