An Asynch-Res Deep Learning Model And Its Application To Road Network Traffic Flow Prediction

Abstract—Intelligent transportation system(ITS)has promoted the development of traffic big data.Among the problems of traffic flow prediction,deep learning methods is suitable for large-scale data sets which have attracted constant attention.Deep learning uses a large number of parameters for nonlinear high-complexity function approximation,which makes deep learning very suitable for handling complex traffic flow prediction problems.Some models of deep learning have already obtained very good predictive performance in road traffic flow prediction problems.However,there are few studies on the traffic forecast of road networks.The current deep learning framework for traffic prediction of road networks generally includes two kinds which are Centralized Modeling and Decentralized Modeling.The prediction ability of Centralized Models cannot fully reflect the ability of deep learning algorithms and the driving force of large-scale datasets.The Decentralized Modelsconstruct simplely and predict accurately,but the extremely high parameter quantities are difficult to applicate in the prediction of large road networks.This paper first compares the deep learning model of traffic flow prediction,modeling,training and model evaluation of MLP,SAE,SimpleRNN,LSTM,and GRU units.Among them,the prediction ability of LSTM and GRU is proved the best with better convergence.In addition,an LSTM or GRU-based Asynch-Res model is proposed to predict road network traffic flow,and the deep learning method of traffic flow prediction is generalized to the road network traffic flow in order to reduce the parameter quantity of the road network traffic flow model.And keep the prediction accuracy close to the original model.According to the analysis of the process and thoughts of ARIMA model for traffic flow data,the features of systematic traffic flow data are divided into obvious and common generalized features and individualized features with not obvious performance,and they exist in the datasets of single or several roads,and proposea method to separate traffic data intogeneralizaed features and individualizedfeatures.In this paper,the structure of the residual network is introduced.The Asynch-Res formula are clearly defined and fully describes the effect of asynchronous residuals,the operating mode and mechanism of the proposed architecture model.This paper uses the Central Module to build a large model for the road network to extract the generalized features of the whole road network,and to provide parameter sharing on each road prediction to increase parameter utilization through parameter reusing,and use Road Modules to learn Asynch-Res sequences for each road to extract road personality characteristics.The Road Module is constructed to improve the prediction accuracy of the model.Then,the merging layer is used to form an Asynch-Res road network traffic flow prediction model.On the UCI’s PEMS-SF machine learning dataset,20 effective Asynch-Res models are built and compared.It is proved that the prediction model of Asynch-Res architecture has the similar accuracy of the original model and the parameter consumption is greatly reduced.Compared with the previous architecture,theAsynch-Resmethod can achieve a prediction accuracy of more than 93%for Acc0.02,and the parameter amount is only about 1/7 of the previous model parameter amount.