Travel Time Forecasting With Combination Of Spatial-temporal And Time Shifting Correlation

With the development of society and the popularization of vehicles and other means of transport,the urban traffic network has become increasingly complex.While traveling,people gradually pay more attention to the time and efficiency of travel.Accurate travel time can not only optimize the travel route,reduce unnecessary travel time for residents,but also reduce traffic flow on busy roads,avoid traffic congestion,and ease traffic pressure.Therefore,the problem of urban short-term travel time prediction has important significance and research value for improving residents' life and urban transportation.It also plays an active role in improving urban transportation and building wisdom cities.At present,the study of short-term travel time prediction is divided into four categories:naive,parametric,non-parametric and hybrid.Naive methods,such as history average(HA),have low computational complexity and are easy of deployment.However,because of the absence of complex calculations,naive methods usually have problems with low accuracy.Parametric methods mean that the model structure has been defined in advance and only the exact values of a given set of parameters need to be determined.These methods are mainly based on time series analysis,including autoregressive moving average(ARMA)model,differential autoregressive moving average(ARIMA)model and spatial-temporal difference autoregressive moving average(STARIMA)model.These methods usually forecast the road travel time in the next time period based on the historical travel time series,and do not consider the spatial characteristics and the influence of other roads.Non-parametric methods refer to the model structure and its parameters need to be determined in training,and are divided into model-based and memory-based.In a model-based approach,historical data is used to build a model structure,and historical data is no longer needed once the model structure is determined.Such methods mainly include Artificial Neural Network(ANN),Random Forest(RF)and Support Vector Machine(SVM).Memories-based methods need to maintain an additional database to store historical data,because historical data is not only used in the construction of the model structure,but also needs to be used in the prediction phase.A typical method is a k-nearest neighbor(kNN)algorithm.In Intelligent Transportation System(ITS),a large amount of two-dimensional urban traffic data can be acquired,and the amount of data is increasing rapidly.At the same time,computer hardware capabilities have been greatly improved.So deep learning model gets more attention recently.The high-dimensional feature extraction capabilities of deep learning methods are extremely important for predicting tasks.In this paper,a hybrid prediction model based on spatial-temporal characteristics of traffic data combined with convolutional neural network(CNN)and long-term memory network(LSTM)is presented.We first use the KL-divergence and urban traffic network to filter the urban roads,and filter out the upstream related roads that have influence on the travel time of the target road.Then a spatial-temporal and time shifting feature matrix is constructed according to the relevant roads.Finally,features are extracted through CNN and travel time prediction is performed using LSTM.The main contributions of this paper can be summarized as follows.We propose a novel deep learning architecture for short-term travel time forecasting,which combines CNN and LSTM neural networks.To the best of our knowledge,it is the first time that the CNN and LSTM are concentrated to predict the short-term travel time;We use KL-divergence and urban road networks to calculate the similarity between roads and then obtain the most top-k relevant roads of target section.Based on the results obtained above,we can greatly reduce the complexity of the input data;We identify a time shifting relationship between different sections.We first extract the information of time shifting feature,and integrate this feature into the existing spatial-temporal methods.We applied the proposed model to the actual traffic data set and conducted a large number of experiments.The experimental results show that our model is better than the current state-of-the-art in terms of accuracy and efficiency.It has practical value and can be applied in related fields.