Traffic Prediction Method Based On Deep Learning

At the historical node when living standards have gradually improved and the task of building a well-off society in an all-round way has been completed,the penetration rate of private cars in China has reached a new high,with an average of two to three families owning one private car.On the other hand,due to the rapid urbanization,traffic infrastructure construction is not perfect,resulting in traffic congestion,traffic accidents and other problems,resulting in huge economic losses and even life safety losses.Under the above background,how to reasonably control the traffic and plan the next infrastructure construction is a major direction of urban development.The concept of smart city,that is,the integration of technology and urban development,puts forward a new direction for urban traffic management and construction.As a part of smart city,traffic prediction method is a major research direction of deep learning.At present,there are many research methods published in the field of transportation,but there are still many problems.First of all,short-term traffic data is prone to mutation,and has strict time requirements.Secondly,the traditional neural network model is not suitable for short-term traffic prediction.Therefore,the traffic prediction method based on deep learning has great value space and research significance.This paper introduces and improves the traffic prediction method based on deep learning(1)A traffic prediction method based on bidirectional gating neural network and attention mechanism is proposed.Considering the strong time correlation of traffic data in a short time,this method constructs a sub interval series.While using bi-directional gating neural network to extract the time correlation characteristics of sub interval series,considering that the previous time series model only uses the output of the last moment as the feature,the research combines the output of the model at different times as the input of the prediction layer,In order to capture the dynamic changes of traffic data in time,the attention mechanism is combined at the output end of the two-way gated neural network to obtain the real distribution of data features and avoid the interference of redundant features on the model.Finally,it is used for model training and achieves good results.(2)On the basis of method 1,this paper first analyzes the periodicity of the traffic data,and verifies the long-term periodicity of the data and the short-term correlation of the data.Then,considering the lack of long-term characteristics of the traffic forecasting method,this paper proposes a method to construct the daily cycle series,In this method,the traffic data of the same time in different days are combined into daily cycle series,and the key features of the series are extracted by using self encoder,which finally solves the defect of insufficient long-term features of the time series model.At the same time,the short-term correlation of the time series is extracted by combining method 1,and the linear and nonlinear features of the data are used for traffic prediction,Compared with the basic models Gru,bigru,and the classic traffic prediction algorithms CNN,SVR,etc.,the RMSE index and MAE index have been improved.