Delay Analysis And Prediction Of Actural Train Performance

In recent years,with the rapid development of China’s urban rail transit construction,the number of urban rail transit lines and operating mileage have continued to increase.As passenger flow continues to increase and train operation intervals continue to shrink,higher requirements are placed on the punctuality of train operation.However,due to various reasons,there will be a delay in the train operation process,affecting the order of urban rail transit.In order to ensure the safe and punctual operation of the train,this paper attempts to analyze the mechanism of train delay and study the distribution of train delay based on the actual performance data collected on site.This paper hopes to propose a method to establish a train delay prediction model,and help dispatchers to make reasonable train operation adjustments by estimating the delay prediction,so as to restore the train operation order as soon as possible,make the train run according to the scheduled plan,and improve work efficiency.The main research contents of this article are as follows:1.Collect the one-month train operation performance data of Suzhou Metro Line 4 as the research object,reconstruct and organize the original data in data format,observe and analyze the actual running track of the train on the line,analyze the reason and mechanism of the train delay,and study the train delay distribution in space and time.2.Establish a feature selection module,exclude some irrelevant features and indistinguishable features through variance filtering,and normalize and uniquely encode different types of variables in feature.The XGBoost algorithm is used to rank the features related to the arrival delay of the next station of the train,and combined with the actual operation of the train to determine the important features to participate in the model training.3.Establish a two-level late prediction model.The first-level prediction model is based on the LSTM neural network algorithm.According to the input features,the existing data is constructed as a 3D tensor,which is the number of samples,the length of the sequence,and the number of features.This 3D tensor is used as the input of the first-level prediction model,and the model output is the arrival time of the next station.The secondary prediction model is a random forest and XGBoost model.The output of the primary prediction model and the output of the middle layer are added to the input of the secondary prediction model to participate in the training of the secondary prediction model.4.Divide the training set and the test set,and find the best parameters of each module in the training set through a combination of five-fold cross-validation,grid search and learning curve.The model is verified and the prediction results are analyzed,and the prediction effects of the model and the fusion model are compared.Finally,the optimal model is selected as the prediction model for the train arrival delay in the next station.