Research On The Prediction And Parallelization Of Urban Area Traffic Flow Based On Deep Learning

With the continuous progress of our country's social economy and the increase of urbanization rate,many large and medium-sized cities have experienced serious road traffic congestion,frequent accidents,and other traffic problems.The increasingly serious urban traffic problems have become one of the important factors restricting the future development of cities.In order to effectively solve these traffic problems in cities,and at the same time effectively promote the construction of smart cities,this paper focuses on the traffic flow prediction problem in urban areas,conducts research on the area division of cities and traffic flow feature extraction,and proposes a prediction model based on deep learning.Moreover,for the purpose of expediting the training process of the predictive model and improving the accuracy of the model,further research has been done on the parallelization of the deep learning model.The main research contents are summarized as follows:Aiming at the problem of urban regional traffic flow prediction,this paper proposes two prediction models.one is a Spatio-temporal residual network model based on dilated convolution.The model uses a convolutional neural network to extract short-distance spatial features and uses a residual network combined with dilated convolution to increase the receptive field to extract long-distance spatial features.It uses a long short-term memory neural network to further model time-series features and uses a fully connected layer to build a model of the impact of external factors such as time and weather on traffic flow.Finally,the extracted Spatio-temporal features and external factor features are merged,and the prediction results of urban regional traffic flow are output.The other is the traffic flow prediction model based on spatiotemporal graph convolution.In this model,we extract the temporal and spatial characteristics of historical time traffic information through the spatiotemporal block composed of temporal convolution network and graph convolution,and output the predicted traffic flow prediction.Aiming at the parallelization of deep learning models,this paper implements three parallel modes for the above-mentioned Spatio-temporal residual network model based on dilated convolution: data parallelism,model parallelism,and gradient accumulation algorithm.In terms of data parallelism,a data slicing algorithm that can maintain the load balance of GPU groups is designed.In terms of model parallelism,the calculation of the training parameters of a single GPU is reduced by dividing the model.Finally,a gradient accumulation algorithm based on deep reinforcement learning is proposed,which automatically adjusts the amount of gradient accumulation in deep learning training,in order to achieve the task of speeding up training process and improving accuracy.In the end,Simulation experiments have been carried out on the public data set of Beijing taxi GPS trajectory.The prediction model was compared with six benchmark model methods.The experiment results demonstrate the forecasting model proposed in this paper has high accuracy in RMSE and MAPE evaluation indexes,with the prediction accuracy increased by 7.1% and 40.6%,respectively.Data parallel mode and gradient accumulation algorithm based on deep reinforcement learning can improve the accuracy and accelerates the convergence of the prediction model,RMSE decreased by 50.12%,MAPE decreased by 59.2%,and convergence speed increased by 36.71%.It shows that the prediction model is suitable for the problem of urban regional traffic flow prediction,and the deep learning model can accelerate the training speed through parallel computing technology.