Data Processing And Combined Model Method For Traffic Flow Prediction

Since the 21 st century,with the improvement of people's living standards,the rate of car ownership has increased significantly in China.According to the statistics from National Bureau of Statistics,Chinese private car ownership in 2018 is 231.22 million,which is still growing.Such a large percentage of private car ownership is often accompanied by a series of problems,the first of all is the congestion problem.At present,due to the rapid increase of vehicles,the public transportation infrastructures in some large cities in China are facing a huge challenge,and the congestion problem has become headache of urban governance.A more effective way to manage traffic problems is to accurately predict traffic flow in advance,and then regulate and divert the road ahead of time.This paper will use the average transit time of roads as an indicator to measure traffic flow,and use data mining to predict the average transit time of roads.The data in this paper is the road data of 132 roads in Guiyang City,Guizhou Province from March to June in 2017,including road attributes,road topology,average transit time per two minutes.The data from March to May was used as the training data.The data in June was used as the testing data,and the average absolute percentage error(MAPE)was used as the evaluation of the model.The aim in this paper is to forecast the average transit time of every two minutes in the three periods,including[8:00,9:00],[15:00,16:00],[18:00,19:00].The main work and innovations of this paper include:(1)Explain related technologies such as traffic flow theory,traffic flow prediction method,and combined model method.(2)Propose a reasonable data cleaning solution for different types of traffic data.Pre-processing includes outlier evaluation,missing value filling,data transformation,data specification.(3)Because the traffic flow data has strong sequence correlation,this paper uses the ARIMA to train the traffic flow data.In addition,considering the complexity of the traffic,this paper also uses XGBoost and LSTM for modeling.XGBoost relies on artificiallyextracting features to improve the effect of the model,while LSTM relies on algorithms to automatically learn features.(4)Because the single model has certain limitations,this paper will ensemble the results to improve the performance of the model.In the process of model fusion,the improved weighted fusion algorithm and stacking algorithm are used.The improved weighted fusion algorithm uses the linear regression model to learn the weighting factors of each single model.In the Stacking process,this paper designs a two-layer classifier.The base classifier includes XGBoost and LSTM.The meta-layer classifier is a linear regression model.