Research On The Ensemble Learning And EAE-FEM Algorithm Based On Transportation Flow Prediction

With the development of the society,transportation has become more and more important and popular.The road conditions are very closely related to the urban congestion,so we should make further research on vehicle traffic flow forecast.And as the most importance part of international trade,the shipping is directly linked to economic development,so it is also worthy of our study.Therefore,this article will study from the shipping traffic and vehicle traffic.The content about shipping in the article is about Estimate Time of Arrival(ETA),which is the predicting time when the shipping arrives at the destination port.It is generally based on historical GPS signals to predict the future status,and finally calculate the ETA.And the modeling and forecasting of the ETA is a difficult problem.We start with analyzing shipping data,and then use the matching algorithm to reconstruct missing data,and use the feature engineering method to construct the data of distance,time,latitude,longitude,speed,direction and their statistics,and finally take these data as input to the prediction model.We study three different types of ensemble machine learning including Bagging,Boosting and Stacking,and we use Random Forest,GBDT and their combination Stacking method to predict the ETA.The experiments prove that Stacking is the best way among the three methods to predict ETA,and also prove that the works of this data processing and feature establishment are beneficial to the accuracy of ETA prediction.The content about vehicle traffic in this paper is the prediction of vehicle speed on the road.Although the prediction of time flow data is well-established in statistics and machine learning,the complex space information in the traffic network makes vehicle traffic flow prediction worthy to research.By studying classic vehicle traffic flow prediction models,we find that each model has its own suitable data.If a model applies in an unsuitable dataset,the performance will be bad.In response to this problem,this paper proposes the EAE-FEM mechanism to mining the hidden information of the time and spatial data.EAE-FEM mechanism consists of three steps.First,the intermediate data is mined by the auto-encoder part.Then,the FEM(Feature Enhancement Mechanism)is introduced to further analyze and process.Finally,the EAE-FEM mechanism and classic model are trained by two-way training.Combining with five state-of-art models of LSTM,GRU,STGCN,TGC_LSTM and GBDT,we apply the EAE-FEM mechanism in two real dataset.The numerous experiments show that EAE-FEM mechanism is effective and it improves the performance of traffic flow prediction.