Traffic Flow Prediction Based On Deep Residual Neural Network

In the field of traffic,how to predict the change of traffic flow accurately is important for urban traffic management.Accurate and real-time traffic flow information can help people make road planning,reduce the likelihood of traffic accidents,make the best use of road resources and relieve traffic pressure.However,due to many factors like bad weather,horrible accidents,predicting time series like traffic flow is full of challenges.At the same time,traffic flow data is time series,so it has the characteristics of non-linearity and periodicity.How to comprehensively consider many factors and accurately predict the traffic flow within a period of future has been discussed thousands of times recently.Firstly,this dissertation analyzes the factors that affect traffic flow prediction in urban areas from many aspects.It proved that for traffic flow there is interaction between time and space in many time periods,and then traditional time series analysis cannot capture such correlation information.In recent years,neural network method is frequently used to solve various problems,compared with other methods,it shows many great advantages.One of the hot networks-Convolutional neural network,can automatically extract the internal correlation characteristics of the inflow data,so it will reduce the errors caused by artificially constructed features.For the problem of traffic flow prediction,at the beginning,we try to use ARIMA model,then find it can fit the trend and periodicity but in some area like the minimum,it does have some error.Then we use LSTM model,which can acquire time dependence from long distance,although this model do really better than ARIMA,it still can improve.Because the deep neural network can better extract the features of the original sequence,the deep residual neural network model is used for the prediction.The experimental results of this model show that,in the case of large amount of data,the root mean square error is smaller than the other models,and the prediction result is better.Therefore,in this dissertation,in order to capture the spatial dependence,we use convolutional neural network to carry out multi-layer convolution operation on the grid graph matrix of the inflow and outflow flows of the urban area,then build model to catch interaction between different areas.For the problem of simultaneously predicting the inflow and outflow flows in an urban area,we propose a deep residual network Re s Net model which composes of convolutional neural network.The model first divides the regions,and then transforms the time series of traffic flow in each small region into a grid diagram sequence with time and space characteristics according to certain rules.Then,grid image segments of flow in three different time ranges were extracted from the two dimensions of inflow flow and outflow flow,which were taken as the input of the model.After that,the same residual network of three parts is designed to capture the closeness,period and trend of traffic data at the same time,and the output of each part is allocated different weights for fusion.At the same time,we design a fully connected neural network to capture the influence of some external factors,such as week attribute,holiday attribute and weather attribute.Finally,aggregating every part of the results to get the final output results.In the end,from "Gaea" which was a plan to provide public data for research,we extract some data form "Chengdu" and by testing these data,the results show that compared with the traditional time series analysis model ARIMA and the neural network model LSTM and other model,Res Net has obvious advantages.It shows that the proposed method can predict regions of urban traffic flow well.