Research On Prediction Model Of Spatio-temporal Data Based On Residual Analysis

With the development of China's economy,people's demand for travel modes is getting higher and higher,and more and more people are more inclined to choose private cars for travel,which has led to urban traffic congestion.If the traffic flow can be predicted in time and the road traffic situation can be grasped,the congested roads can be dredged in advance to improve the mobility of the city.This has a very important practical significance for both traffic managers and traffic participants significance.However,in the process of forecasting actual traffic flow,there are often two difficult problems:(1)The data of traffic flow changes variously,and the performance in both the time dimension and the space dimension varies widely,so the feature extraction of traffic flow has become a major technical difficulty in traffic flow prediction.(2)Due to the difficulty in extracting features,many existing models increase the accuracy of experiments by increasing the depth of the network,which can easily lead to problems such as gradient disappearance,gradient explosion,and network degradation,and the experimental results are not ideal.In order to solve the above problems,this paper proposes a prediction model of spatio-temporal data based on residual analysis(RAM-TF)for traffic flow prediction.In order to better extract the spatial and temporal characteristics of spatio-temporal data,the spatial-temporal data is classified and collected into three categories: proximity spatio-temporal data,trending spatio-temporal and periodic spatio-temporal data.The data are processed with a convolutional neural network(CNN)to extract the spatial features of the spatiotemporal dataset.After preliminary convolution,the spatio-temporal data is fused by matrix addition,which facilitates the centralized training and optimization of the entire model,while simplifying the model.The model is optimized using a residual network(Res Net),where the residual network is composed of residual units in series.The residual unit structure used in this paper consists of convolution operations,Relu functions,and layer-hopping links.In order to make the prediction effect closer to the real situation,we have also added noise(such as rain,holidays,etc.)to predict the future traffic flow.The noise processing includes two fully connected and two Relu functions.In general,the specific contributions of this article are as follows:(1)Spatio-temporal data is divided into three parts: proximity spatio-temporal data,periodic spatio-temporal data,trending spatio-temporal data to deal with the temporal characteristics of spatio-temporal data;Convolutional Neural Network(CNN)is used to extract the spatial characteristics of spatio-temporal data.(2)Res Net is used to optimize the model: The residual network can deal with the depth of the network in a targeted manner.The residual network used in this paper is a series of residual units,where the residual units includes convolution(CNN),Relu function,and layer-hopping links,both for feature extraction and optimization model.(3)Fusion: In order to enable the model to be concentratedly trained and optimized,the model uses matrix addition to fuse the spatio-temporal data(Fusion),while simplifying the model.(4)In this paper,experiments are performed on the number of units included in the residual network of 3,4,5,and 6,respectively.The experimental data used are Taxi BJ data in Beijing and Bike NYC in New York.We compare the proposed model with five existing models on the test set to prove the superiority of the RAM-TF model.