Combining Graph Convolutions And Recurrent Neural Networks For Traffic State Prediction

Traffic prediction is critical in intelligent transportation systems(ITS).The modern city is evolving into a smart city on a gradual basis.Accurate traffic prediction will aid in route planning,vehicle dispatching,congestion relief,traffic management measures,and road infrastructure design.Traffic prediction’s objective is to predict the potential traffic states in the traffic network given a sequence of historical traffic states and the physical roadway network.Traffic prediction has recently gained increased attention as in result of the popularity of the online navigation systems,ridesharing,and smart city initiatives.Since,the road traffic states are non-stationary,prediction accuracy is a fundamental problem and subject of the study.Due to the time-varying nature of traffic patterns and the complex spatial dependencies of road networks,traffic prediction is an especially difficult application of spatiotemporal data prediction.In addition to the spatial position on the map,the list of features includes the date,time,holiday or working days,and special conditions.Environmental factors such as weather patterns,road quality,incidents,and road characteristics all have an impact on traffic congestion.The major challenges must be resolved by predicting short-term traffic predictions,providing of long-term traffic predictions,congestion,travel times,traffic management insights,preventing traffic problems due to weather,road accidents,etc.The main problems are how the complex and dynamic spatial and time dependencies can be model and how accurately predicted time-changing patterns of transport and complex spatial dependencies on road networks etc.can be achieve.While traffic prediction have made considerable progress in recent years,still many open problems have not completely explored.These problems must be resolved now.Transport’s future will largely rely on the best realistic traffic prediction.The problem for exact prediction,currently,arises from how spatial temporal interaction could be capture in traffic data,how external factor effect on road states can be modelled,how short-term and long-term predictions can accurately predicted simultaneously,etc.Many study-methods have developed in less complex traffic systems and relatively limited sizes of traffic data sets,focused on traffic prediction using statistical methods.However,statistical models are very limited in their capacity to handle high-dimensional time series data.Existing models such as ARMA or ARIMA are essentially linear and cannot explain traffic states’ stochastic or non-linear existence.In addition,it is always impossible and difficult to view the network parameters in terms of actual spatial dependencies.Rapid advancement in computing capacity and traffic data volume growth have been based on deep traffic prediction learning techniques using CNN,GNN,GCN,RNN,ASTGNN etc.Deep learning models for traffic predictions are capable of effectively learning high-dimensional features and producing the accurate predictions.Additionally,recurrent neural networks(RNNs)and their versions,such as long short-term memory(LSTM),have shown considerable promise for solving traffic prediction problems.Due to the fact,that LSTM and its variants are effective at capturing the long-temporal dependencies inherent in sequential data,they have been widely applied in a variety of sectors,including transportation,energy,medicine,and economics.To answer this challenge and to propose a solution,the thesis studied the traffic network and proposed a novel deep learning model called "Deep-Graph Recurrent Network(Deep-GRN)" that utilizes Graph Convolutional and Recurrent Neural Networks to comprehend and optimize the interactions between roadways in the traffic network,as well as predicts network-wide traffic state.After studying the complex spatial and dynamic temporal dependencies inherent in traffic data,the thesis used real-world traffic data and graph convolution networks to model the spatial associations of road network structure.The thesis also used RNN-based neural networks to model the temporal dependencies inherent in traffic time series.Deep learning models finds significantly more features and complex architectures than traditional approaches,hence it is enable to achieve higher efficiency.In this work,by using my proposed model,the thesis predicted the state of a certain road section at a certain time in the future.Let call this section “link”.In order to predict the accurate traffic state,the thesis build three models to extract useful information from the three types of data.These models were NFM model(Bi-interaction,MLP),LSTM(pooling techniques)and GCN(topology).The thesis obtain results on Di Di(Taxi)datasets of one month of collection of data,consisting on these three following sub-datasets,traffic,topo and attr,to test and verify the efficiency of the model and experiment.Di Di dataset provides the academic community with real and free data desensitization tools.First,there are several category data for every link itself,including the connection level,speed limit level,etc.The data for these categories reflect each link’s characteristics.For prediction of traffic conditions,the features of the link itself are very significant.To do this,the category data is been generated by NFM model(neural factorization machines).Furthermore,a certain traffic states are linked closely to the traffic status of a previous time and to the same historical period.The thesis therefore need to extract data and historical data from recent data.These are sequence-shaped data,or called“sequence data”.The most commonly used and efficient model for processing sequence data is Recurrent Neural Network.The thesis use LSTM to extract data from the two separate time sequence data.Long Short-Term Memory(LSTM)networks were the most successful solution for almost all these sequence prediction problems.The increased expressive power of LSTMs will result in better results if you have sufficient data.The roads are linked which is important.The state of each link is also important from where that link is connection.Considering each link as node,the thesis build a linked diagram in a certain area of all links.Using GCN models,the thesis discovered how the linked graph influences each other.The thesis used GCN to extract information from the subgraph created by it and its surrounding nodes that was very helpful for my future work when the thesis was about to predict the future status of a connection.After performing these evaluation and feature extraction tasks on the Di Di dataset,the thesis provides estimation and analysis by showing comprehensive experiments for comparing the performance of various approaches on my real-world public dataset Di Di.The thesis situates these three kinds of useful information together and used DNN and LGB model alongside with my proposed model,Deep-GRN,for analyzing results and for predict the state of link in the future.Finally,my aim achieved and the thesis got significant prodigious results.As by results,my model is significantly better and accurate than the other two baselines,DNN and LGB.Results clearly show that my model is more efficient and accurate.After accomplishing the computational goal,the thesis successfully predicted the state of a certain road section at a certain time in the future.Moreover,my results are far better few steps ahead than the other two baselines,DNN and LGB model.It is worth mentioning that the thesis predicted the state of a certain road section at a certain time in the future.Pointing out and analyzing the three core metric,category data(generated by NFM model),sequence data(generated by LSTM)and topology data,DNN and LGB only takes input from category data and sequence data and they cannot process topology data.In addition,both of these are not as good as my selected models for category data and sequence data.My model strongly considers topology data and it performs far well in computing category and sequence data.The explanation for my improved performance and results is that the thesis used a variety of prodigious methods to discover possible solutions to the prediction problem,including the use of powerful evaluation functions and models such as NFM,LSTM,and GCN.In related prediction works,LSTM is frequently used.Researchers have constructed notable spatio-temporal deep learning architectures using a combination of CNNs,LSTMs,and graph neural networks(GCNs).However,issues such as complex data preprocessing and input constraints(e.g.,adjacent matrix)make these models difficult to handle.However,in my proposed model,Deep-GRN,the thesis concentrated on adjacent matrix extraction during topo feature extraction.Moreover,the most important feature of my model is that it uses topological information(GCN),while the baseline methods does not.The introduction of topology information makes each link have its own and other link’s information at the same time.I have used different models(NFM,LSTM and GCN)for different data(category data,sequence data,and topology data),rather than connecting all the data together and use one model.During experiment,if GCN will removed from the model,it does not perform well enough.Without GCN part,model could well trained in shorter time but loss value cannot reach a lower level,which means the model is not well enough in performance.The introduction of GCN adds more parameters to the model,which leads to more time training,but the improvement of the effect is significant and worthwhile.At the same time,it is worth noting that the introduction of more parameters only increases the number of training epochs,and the time cost in predicting a single epoch of data is negligible.My contribution is advantageous in both economics and academia.In the economic realm,effective traffic state prediction has the ability to increase traffic conditions and decrease delays by promoting more efficient use of available functions such as traffic management strategies,road infrastructure settings etc.The transportation industry has produced a large amount of data in the age of deep learning and big data.We can predict traffic conditions using many different data sets based on this data.It can improve the user experience when it comes to understanding traffic predictions using deep learning models.In the academic context,this study will aid in the understanding of the complex and dynamic spatial and temporal relationships inherent in traffic data with straightforward explanations.